Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-MTP-GGUF

ABSOLUTE FIRE: 640 ARC-C for both 8bit and 4bit. Model exceeds 7 of 7 benchmarks for Qwen 3.5 9B, Qwen3.5 27B, Qwen3.6 35B-A3B, and meets Qwen 3.6 27B in some cases... and it does so in 4bit and 8bit. Regular and MTP (fast) NEO IMATRIX GGUFs provided. (this model is part of the Qwen 3.6 27B Fable Fusion 711 pipelines: 2200+ likes, 3 million + downloads)

NEW - Qwen 3.8 Reasoning Modes: 2 MTP quants (Q6/Q8) Now with 5 reasoning modes (2 new - Spoon / Einstein), and 5 instruct modes (2 new - Spoon / Einstein, all use ZERO REASONING TOKENS) all switchable on the fly via API, direct and "in chat" (yes - model ctrl at the chat/message level). Model name has "plusIQ" in the name. (there is also a extra robust "tools" version too.)

Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-MTP-GGUF

Extreme intelligence in a small package. Jaw dropping performance. Superior instruction following.

A multi-stage and multi-model fine tune and multi-stage merge on local hardware by myself and Nightmedia.

Several of my 9B Qwen 3.5 fine tunes were used to make this model.

The goal of this model was to raise the general intelligence of the model AND instruction following. Getting these right raises the performance of the model from top to bottom.

Keep in mind, even thou this model EXCEEDS 7 critical benchmarks of Qwen 3.5 27B model, that does not mean it will beat the 27B in all tasks.

This is also a HERETIC model, trained post "Heretic'ing" -> this model does what you want, no questions asked.

Also: Thinking/reasoning block has been compacted, and in many cases it is a lot stronger.

Fully uncensored.

Benchmarks below.

Example generations at the bottom of the page.

PS:

If you are looking for a 27B model that BEATS Qwen 3.6 27B, Qwen 3.5 27B, and Qwen 35B-A3B, IN ALL critical benchmarks this is the one:

https://huggingface.co/DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF

BENCHMARKS: (Bar Graph below.)

          arc/c  arc/e boolq hswag obkqa piqa  wino

Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic [instruct]
bf16      0.649,0.832,0.895,0.713,0.482,0.783,0.699
mxfp8     0.647,0.836,0.895,0.706,0.460,0.784,0.695
mxfp4     0.640,0.824,0.886,0.703,0.468,0.780,0.691

Qwen3.5-9B-Instruct: [base, non heretic]
mxfp8     0.571,0.719,0.895,0.683,0.426,0.770,0.671

---

Qwen3.8-27B:  [base, non heretic]
mxfp8     0.591,0.782,0.896,0.746,0.448,0.801,0.711
mxfp4     0.581,0.771,0.889,0.738,0.442,0.798,0.713

Qwen3.6-27B-Instruct: [base, non heretic]
mxfp8     0.647,0.803,0.910,0.773,0.450,0.806,0.742

Qwen3.5-27B-Instruct: [base, non heretic]
mxfp8     0.557,0.711,0.868,0.533,0.452,0.706,0.695

Qwen3.6-35B-A3B-Instruct [base, non heretic]
mxfp8     0.581,0.757,0.892,0.751,0.428,0.803,0.688

Graph note: First three bars are Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic.

"Bench" NOTES: - Models are tested in "Instruct" mode because this generally works better with the testing harness. - Testing via "thinking" mode also shows the metrics (and changes) but not the true extent. - In actual fact when the model IS in thinking mode, it will exceed INSTRUCT benchmark scores in most cases.

Regular and MTP GGUFS:

All quants (regular and MTP) are NEO IMATRIX, which improve accuracy of the quants by an additional 2-4% over normal GGUFs as well as long context performance.

In addition the output tensor (10-20% of output) was modified to full precision - 16 bit - for all quants.

"MTP" GGUFS (multi-token prediction): - "MTP" GGUFS will have "MTP" in the name as a suffix. - I have also set the MTP tensors to Q8_0 precision for all quants. - To get better performance keep temp 1 or less (higher temps degrade MTP performance). - Likewise with rep pen ; keep at 1 (off). If you raise it performance will suffer. - If you see "token acceptance" rates BELOW 50% (predict 2 tokens) switch to normal quants.

SPEED: - On Q4_K_S (4bit) quant, regular GGUFs are about 130 t/s, whereas MTP GGUFs (acceptance at 60%, 2 tokens) can exceed 185 T/S. (5090, Windows 11, testing in LMStudio) - Speeds will vary depending on GPU(s), AI app, O/S (Linux/Mac will generally be faster) and hardware. - "MTP" quants speeds will vary ; for creative/complex and/or temps over 1 use regular GGUFs for better performance.

I suggest you download at least one of each - regular and MTP gguf(s) - and test them for your use case(s).

If you get "token acceptance" (predict 2 tokens) with MTP quant(s) BELOW 50% (this means regular quants will run faster), then regular GGUF(s) will actually perform better - ie faster.

MTP quant(s) can in some cases run faster as the token window fills up and/or in multi turn chats.

Note there is NO other diffence between the quants type besides speed: both will do the same job.

Model: - 256k context - Gguf quants run in all standard AI apps. - Vision is activated, but you need to download separate "mmproj" file (ONE) to use it.

VISION: - Vision (images) tested. - You need an "mmproj" (just one) of these downloaded too, and placed in the same folder as the GGUF for images.

Qwen Model Settings (suggested):

• Thinking mode for general tasks: temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0
• Thinking mode for precise coding tasks (e.g. WebDev): temperature=0.6, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0
• Instruct (or non-thinking) mode: temperature=0.7, top_p=0.80, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
• Context window min from 8k to 16k.

IMPORTANT - 5 Reasoning modes and 5 instruct modes:

GGUFS: - Model name has "plusIQ" in the name. (there is also a extra robust "tools" version too with "plusIQ-TOOLS" in the name.) - Avail in Q6 and Q8 MTP.

The good news is this:

All the defaults are still the same for this model, that is "reasoning" is set at "xhigh" and if you activate "instruct mode" it will set automatically at "medium".

This was done to ensure "drop in" of this model into your workflow would work without issues/adjustments.

The GREAT NEWS is this:

You are no longer limited to these defaults, and the both new reasoning modes and all modes of instruct are also unlocked too.

Previously if you used "instruct" mode (thinking off) you were limited to only "medium" power with the model.

We modified it so you now have "xhigh", "medium" and "low" available too (as well as 2 new modes - more on that in a minute)

We also modified the model so you can easily access and change between multiple reasoning and instruct modes at ANY TIME.

Here are the reasoning modes: - (There is no) "spoon" -> ULTRA xhigh, research mode // hyper detailed; this will automatically use more reasoning tokens (reasoning mode). - "einstein" -> a "high" mode that spawns up to 20 virtual agents to solve tasks; this will automatically use more reasoning tokens (reasoning mode). - "xhigh", "medium" and "low" -> Standard Qwen 3.8 reasoning modes.

All these modes are also available via "instruct mode" too.

SPECIAL NOTE:

"medium" is the default (both reasoning and instruct) for the model - IE no enhancement.

How strong are the enhancements?

See 22k output here in this PDF, using Q6_K in "spoon" mode:

Spoon Reasoning example PDF

Next, we added coding to all "in message switching" (RIGHT IN CHAT) :

{REASON:xxx} => Where "xxx" is "spoon", "einstein", "xhigh", "medium" and "low".

For instruct mode, just add an "i":

{REASON:xxx} => Where "xxx" is "ispoon", "ieinstein", "ixhigh", "imedium" and "ilow".

THE LAST "reason" / "instruct" mode will be the one used until you switch it again IN THE CURRENT CHAT.

EXAMPLES: - {REASON:einstein} tell me a story. - {REASON:ispoon} tell me a story.

Note the "{REASONxxx}" can be anywhere in the prompt, and will PERSIST until you change it again in the CHAT WINDOW/CURRENT CHAT.

If you open a new chat window (depending on your AI APP) the DEFAULT reasoning mode at the default setting will take over unless you use the "{REASONxxx}" in the new prompt(s) at least ONCE in the new chat window/new chat session.

Also the systems automatically remove it from the "message stream" so the generation is "pure".

For API this can be set manually - "reasoning" (you can use KW args to set this):

reasoning_effort = 'xhigh'
enable_thinking = 'true'

For API this can be set manually - "instruct" (you can use KW args to set this):

reasoning_effort = 'ixhigh'
enable_thinking = 'false'

ADVANCED:

You can modify the jinja template to CHANGE the defaults by using the API code noted above.

Place the defaults you want at the TOP of the jinja template.

---

DE-CENSORING:

Performance

KLD of less than 1 is excellent, zero is perfect.

NOTES: - Suggest min q4ks (non-imatrix) or IQ3S (imatrix). - Tested with rep pen of 1 (off). - Context: 256k (default).

IMPORTANT: - Other versions in testing. - Information from Qwen's repo below. - Video portions of the model were NOT TESTED.

---

The SUPER Qwen Universe - 40B, 27B and 9B ; meet the performance trendsetters:

---

Qwen3.6 27B: The strongest, overall qwen ever beating all other Qwens in total operational power with over 2300 likes // 4 million+ total downloads: - https://huggingface.co/DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF

Qwen3.8 27B: The highest scoring Qwen in brute, raw intelligence, using Qwen 3.8's 3 new reasoning modes, plus token reduction (1/2 to 1/10) enhancements: - https://huggingface.co/DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NEO-CODER-MAX-MTP-GGUF

Qwen3.8 27B: Super smart and 1/2 to 1/20 the reasoning tokens AND 5 reasoning/5 instruct modes switchable on the fly (even in chat): - https://huggingface.co/DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-NM-DAU-NEO-MTP-GGUF

Qwen3.8 27B: 99% power of BF 16 at 4 and 8 bit. Power, Control and NO DE censoring for ultimate performance also with reasoning token reductions: - https://huggingface.co/DavidAU/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-NM-DAU-NEO-MAX-MTP-GGUF

Qwen3.6 40B: The 40B Monster, specializing in creative and research with 730+ likes and over 2 million downloads: - https://huggingface.co/DavidAU/Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking-NEO-CODE-Di-IMatrix-MAX-GGUF

Qwen3.5 9B: At just 9B parameters it beats most untuned 27B models in both intelligence (640 ARC-C) and performance, plus features 5 reasoning and 5 instruct modes (Qwen 3.8) too: - https://huggingface.co/DavidAU/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-MTP-GGUF

---

Using an "uncensored" (refusals removed) model VS trained "uncensored" model

Usually when you a tell a model to generate horror, swear or x-rated content this is all you have to do to get said content type.

In the case of this model, it will not refuse your request, however it needs to be "pushed" a bit / directed a bit more in SOME CASES.

Although this model will generated x-rated content too, likewise you need to tell it to use "slang" (and include the terms you want) to get it generate the content correctly as the "expected" content level too.

Without these added directive(s), the content can be "bland" by comparison to an "uncensored model" or model trained on uncensored content.

Roughly, the model tries to generate the content but the "default" setting(s) are so "tame" it needs a push to generate at expected graphic, cursing or explicit levels.

Even with minimal direction (ie, use these words to swear: x,y,z), this will be enough to push the model to generate the requested content in the ahh... expected format.

---

Settings: CHAT / ROLEPLAY and/or SMOOTHER operation of this model:

In "KoboldCpp" or "oobabooga/text-generation-webui" or "Silly Tavern" ;

Set the "Smoothing_factor" to 1.5

: in KoboldCpp -> Settings->Samplers->Advanced-> "Smooth_F"

: in text-generation-webui -> parameters -> lower right.

: In Silly Tavern this is called: "Smoothing"

NOTE: For "text-generation-webui"

-> if using GGUFs you need to use "llama_HF" (which involves downloading some config files from the SOURCE version of this model)

Source versions (and config files) of my models are here:

https://huggingface.co/collections/DavidAU/d-au-source-files-for-gguf-exl2-awq-gptq-hqq-etc-etc-66b55cb8ba25f914cbf210be

OTHER OPTIONS:

• Increase rep pen to 1.1 to 1.15 (you don't need to do this if you use "smoothing_factor")

• If the interface/program you are using to run AI MODELS supports "Quadratic Sampling" ("smoothing") just make the adjustment as noted.

Highest Quality Settings / Optimal Operation Guide / Parameters and Samplers

This a "Class 1" model:

For all settings used for this model (including specifics for its "class"), including example generation(s) and for advanced settings guide (which many times addresses any model issue(s)), including methods to improve model performance for all use case(s) as well as chat, roleplay and other use case(s) please see:

[ https://huggingface.co/DavidAU/Maximizing-Model-Performance-All-Quants-Types-And-Full-Precision-by-Samplers_Parameters ]

You can see all parameters used for generation, in addition to advanced parameters and samplers to get the most out of this model here:

[ https://huggingface.co/DavidAU/Maximizing-Model-Performance-All-Quants-Types-And-Full-Precision-by-Samplers_Parameters ]

---

Qwen3.5-9B

[!Note] This repository contains model weights and configuration files for the post-trained model in the Hugging Face Transformers format.

These artifacts are compatible with Hugging Face Transformers, vLLM, SGLang, KTransformers, etc.

Over recent months, we have intensified our focus on developing foundation models that deliver exceptional utility and performance. Qwen3.5 represents a significant leap forward, integrating breakthroughs in multimodal learning, architectural efficiency, reinforcement learning scale, and global accessibility to empower developers and enterprises with unprecedented capability and efficiency.

Qwen3.5 Highlights

Qwen3.5 features the following enhancement:

• Unified Vision-Language Foundation: Early fusion training on multimodal tokens achieves cross-generational parity with Qwen3 and outperforms Qwen3-VL models across reasoning, coding, agents, and visual understanding benchmarks.

• Efficient Hybrid Architecture: Gated Delta Networks combined with sparse Mixture-of-Experts deliver high-throughput inference with minimal latency and cost overhead.

• Scalable RL Generalization: Reinforcement learning scaled across million-agent environments with progressively complex task distributions for robust real-world adaptability.

• Global Linguistic Coverage: Expanded support to 201 languages and dialects, enabling inclusive, worldwide deployment with nuanced cultural and regional understanding.

• Next-Generation Training Infrastructure: Near-100% multimodal training efficiency compared to text-only training and asynchronous RL frameworks supporting massive-scale agent scaffolds and environment orchestration.

For more details, please refer to our blog post Qwen3.5.

Model Overview

• Type: Causal Language Model with Vision Encoder
• Training Stage: Pre-training & Post-training
• Language Model
  • Number of Parameters: 9B
  • Hidden Dimension: 4096
  • Token Embedding: 248320 (Padded)
  • Number of Layers: 32
  • Hidden Layout: 8 × (3 × (Gated DeltaNet → FFN) → 1 × (Gated Attention → FFN))
  • Gated DeltaNet:
    • Number of Linear Attention Heads: 32 for V and 16 for QK
    • Head Dimension: 128
  • Gated Attention:
    • Number of Attention Heads: 16 for Q and 4 for KV
    • Head Dimension: 256
    • Rotary Position Embedding Dimension: 64
  • Feed Forward Network:
    • Intermediate Dimension: 12288
  • LM Output: 248320 (Padded)
  • MTP: trained with multi-steps
• Context Length: 262,144 natively and extensible up to 1,010,000 tokens.

Benchmark Results

Language

Vision Language

Quickstart

[!Important] Qwen3.5 models operate in thinking mode by default, generating thinking content signified by <think>\n...</think>\n\n before producing the final responses. To disable thinking content and obtain direct response, refer to the examples here.

For streamlined integration, we recommend using Qwen3.5 via APIs. Below is a guide to use Qwen3.5 via OpenAI-compatible API.

Serving Qwen3.5

Qwen3.5 can be served via APIs with popular inference frameworks. In the following, we show example commands to launch OpenAI-Compatible API servers for Qwen3.5 models.

[!Important] Inference efficiency and throughput vary significantly across frameworks. We recommend using the latest framework versions to ensure optimal performance and compatibility. For production workloads or high-throughput scenarios, dedicated serving engines such as SGLang, KTransformers or vLLM are strongly recommended.

[!Important] The model has a default context length of 262,144 tokens. If you encounter out-of-memory (OOM) errors, consider reducing the context window. However, because Qwen3.5 leverages extended context for complex tasks, we advise maintaining a context length of at least 128K tokens to preserve thinking capabilities.

SGLang

SGLang is a fast serving framework for large language models and vision language models. SGLang from the main branch of the open-source repository is required for Qwen3.5, which can be installed using the following command in a fresh environment:

uv pip install 'git+https://github.com/sgl-project/sglang.git#subdirectory=python&egg=sglang[all]'

See its documentation for more details.

The following will create API endpoints at http://localhost:8000/v1:

• Standard Version: The following command can be used to create an API endpoint with maximum context length 262,144 tokens using tensor parallel on 8 GPUs.

  shell python -m sglang.launch_server --model-path Qwen/Qwen3.5-9B --port 8000 --tp-size 1 --mem-fraction-static 0.8 --context-length 262144 --reasoning-parser qwen3

• Tool Use: To support tool use, you can use the following command.

  shell python -m sglang.launch_server --model-path Qwen/Qwen3.5-9B --port 8000 --tp-size 1 --mem-fraction-static 0.8 --context-length 262144 --reasoning-parser qwen3 --tool-call-parser qwen3_coder

• Multi-Token Prediction (MTP): The following command is recommended for MTP:

  shell python -m sglang.launch_server --model-path Qwen/Qwen3.5-9B --port 8000 --tp-size 1 --mem-fraction-static 0.8 --context-length 262144 --reasoning-parser qwen3 --speculative-algo NEXTN --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4

vLLM

vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs. vLLM from the main branch of the open-source repository is required for Qwen3.5, which can be installed using the following command in a fresh environment:

uv pip install vllm --torch-backend=auto --extra-index-url https://wheels.vllm.ai/nightly

See its documentation for more details.

For detailed Qwen3.5 usage guide, see the vLLM Qwen3.5 recipe.

The following will create API endpoints at http://localhost:8000/v1:

• Standard Version: The following command can be used to create an API endpoint with maximum context length 262,144 tokens using tensor parallel on 8 GPUs.

  shell vllm serve Qwen/Qwen3.5-9B --port 8000 --tensor-parallel-size 1 --max-model-len 262144 --reasoning-parser qwen3

• Tool Call: To support tool use, you can use the following command.

  shell vllm serve Qwen/Qwen3.5-9B --port 8000 --tensor-parallel-size 1 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder

• Multi-Token Prediction (MTP): The following command is recommended for MTP:

  shell vllm serve Qwen/Qwen3.5-9B --port 8000 --tensor-parallel-size 1 --max-model-len 262144 --reasoning-parser qwen3 --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":2}'

• Text-Only: The following command skips the vision encoder and multimodal profiling to free up memory for additional KV cache:

  shell vllm serve Qwen/Qwen3.5-9B --port 8000 --tensor-parallel-size 1 --max-model-len 262144 --reasoning-parser qwen3 --language-model-only

KTransformers

KTransformers is a flexible framework for experiencing cutting-edge LLM inference optimizations with CPU-GPU heterogeneous computing. For running Qwen3.5 with KTransformers, see the KTransformers Deployment Guide.

Hugging Face Transformers

Hugging Face Transformers contains a lightweight server which can be used for quick testing and moderate load deployment. The latest transformers is required for Qwen3.5:

pip install "transformers[serving] @ git+https://github.com/huggingface/transformers.git@main"

See its documentation for more details. Please also make sure torchvision and pillow are installed.

Then, run transformers serve to launch a server with API endpoints at http://localhost:8000/v1; it will place the model on accelerators if available:

transformers serve --force-model Qwen/Qwen3.5-9B --port 8000 --continuous-batching

Using Qwen3.5 via the Chat Completions API

The chat completions API is accessible via standard HTTP requests or OpenAI SDKs. Here, we show examples using the OpenAI Python SDK.

Before starting, make sure it is installed and the API key and the API base URL is configured, e.g.:

pip install -U openai

# Set the following accordingly
export OPENAI_BASE_URL="http://localhost:8000/v1"
export OPENAI_API_KEY="EMPTY"

[!Tip] We recommend using the following set of sampling parameters for generation - Thinking mode for general tasks: temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0 - Thinking mode for precise coding tasks (e.g. WebDev): temperature=0.6, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0 - Instruct (or non-thinking) mode for general tasks: temperature=0.7, top_p=0.8, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0 - Instruct (or non-thinking) mode for reasoning tasks: temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0

Please note that the support for sampling parameters varies according to inference frameworks.

Text-Only Input

from openai import OpenAI
# Configured by environment variables
client = OpenAI()

messages = [
    {"role": "user", "content": "Type \"I love Qwen3.5\" backwards"},
]

chat_response = client.chat.completions.create(
    model="Qwen/Qwen3.5-9B",
    messages=messages,
    max_tokens=81920,
    temperature=1.0,
    top_p=0.95,
    presence_penalty=1.5,
    extra_body={
        "top_k": 20,
    }, 
)
print("Chat response:", chat_response)

Image Input

from openai import OpenAI
# Configured by environment variables
client = OpenAI()

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image_url",
                "image_url": {
                    "url": "https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.5/demo/CI_Demo/mathv-1327.jpg"
                }
            },
            {
                "type": "text",
                "text": "The centres of the four illustrated circles are in the corners of the square. The two big circles touch each other and also the two little circles. With which factor do you have to multiply the radii of the little circles to obtain the radius of the big circles?\nChoices:\n(A) $\\frac{2}{9}$\n(B) $\\sqrt{5}$\n(C) $0.8 \\cdot \\pi$\n(D) 2.5\n(E) $1+\\sqrt{2}$"
            }
        ]
    }
]

chat_response = client.chat.completions.create(
    model="Qwen/Qwen3.5-9B",
    messages=messages,
    max_tokens=81920,
    temperature=1.0,
    top_p=0.95,
    presence_penalty=1.5,
    extra_body={
        "top_k": 20,
    }, 
)
print("Chat response:", chat_response)

Video Input

from openai import OpenAI
# Configured by environment variables
client = OpenAI()

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "video_url",
                "video_url": {
                    "url": "https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.5/demo/video/N1cdUjctpG8.mp4"
                }
            },
            {
                "type": "text",
                "text": "Summarize the video content."
            }
        ]
    }
]

# When vLLM is launched with `--media-io-kwargs '{"video": {"num_frames": -1}}'`,
# video frame sampling can be configured via `extra_body` (e.g., by setting `fps`).
# This feature is currently supported only in vLLM.
#
# By default, `fps=2` and `do_sample_frames=True`.
# With `do_sample_frames=True`, you can customize the `fps` value to set your desired video sampling rate.
chat_response = client.chat.completions.create(
    model="Qwen/Qwen3.5-9B",
    messages=messages,
    max_tokens=81920,
    temperature=1.0,
    top_p=0.95,
    presence_penalty=1.5,
    extra_body={
        "top_k": 20,
        "mm_processor_kwargs": {"fps": 2, "do_sample_frames": True},
    }, 
)

print("Chat response:", chat_response)

Instruct (or Non-Thinking) Mode

[!Important] Qwen3.5 does not officially support the soft switch of Qwen3, i.e., /think and /nothink.

Qwen3.5 will think by default before response. You can obtain direct response from the model without thinking by configuring the API parameters. For example,

from openai import OpenAI
# Configured by environment variables
client = OpenAI()

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image_url",
                "image_url": {
                    "url": "https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.5/demo/RealWorld/RealWorld-04.png"
                }
            },
            {
                "type": "text",
                "text": "Where is this?"
            }
        ]
    }
]

chat_response = client.chat.completions.create(
    model="Qwen/Qwen3.5-9B",
    messages=messages,
    max_tokens=32768,
    temperature=0.7,
    top_p=0.8,
    presence_penalty=1.5,
    extra_body={
        "top_k": 20,
        "chat_template_kwargs": {"enable_thinking": False},
    }, 
)
print("Chat response:", chat_response)

[!Note] If you are using APIs from Alibaba Cloud Model Studio, in addition to changing model, please use "enable_thinking": False instead of "chat_template_kwargs": {"enable_thinking": False}.

Agentic Usage

Qwen3.5 excels in tool calling capabilities.

Qwen-Agent

We recommend using Qwen-Agent to quickly build Agent applications with Qwen3.5.

To define the available tools, you can use the MCP configuration file, use the integrated tool of Qwen-Agent, or integrate other tools by yourself.

import os
from qwen_agent.agents import Assistant

# Define LLM
# Using Alibaba Cloud Model Studio
llm_cfg = {
    # Use the OpenAI-compatible model service provided by DashScope:
    'model': 'Qwen3.5-9B',
    'model_type': 'qwenvl_oai',
    'model_server': 'https://dashscope.aliyuncs.com/compatible-mode/v1',
    'api_key': os.getenv('DASHSCOPE_API_KEY'),

    'generate_cfg': {
        'use_raw_api': True,
        # When using Dash Scope OAI API, pass the parameter of whether to enable thinking mode in this way
        'extra_body': {
            'enable_thinking': True
        },
    },
}

# Using OpenAI-compatible API endpoint.
# functionality of the deployment frameworks and let Qwen-Agent automate the related operations.
#
# llm_cfg = {
#     # Use your own model service compatible with OpenAI API by vLLM/SGLang:
#     'model': 'Qwen/Qwen3.5-9B',
#     'model_type': 'qwenvl_oai',
#     'model_server': 'http://localhost:8000/v1',  # api_base
#     'api_key': 'EMPTY',
#
#     'generate_cfg': {
#         'use_raw_api': True,
#         # When using vLLM/SGLang OAI API, pass the parameter of whether to enable thinking mode in this way
#         'extra_body': {
#             'chat_template_kwargs': {'enable_thinking': True}
#         },
#     },
# }

# Define Tools
tools = [
    {'mcpServers': {  # You can specify the MCP configuration file
            "filesystem": {
                "command": "npx",
                "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/xxxx/Desktop"]
            }
        }
    }
]

# Define Agent
bot = Assistant(llm=llm_cfg, function_list=tools)

# Streaming generation
messages = [{'role': 'user', 'content': 'Help me organize my desktop.'}]
for responses in bot.run(messages=messages):
    pass
print(responses)

# Streaming generation
messages = [{'role': 'user', 'content': 'Develop a dog website and save it on the desktop'}]
for responses in bot.run(messages=messages):
    pass
print(responses)

Qwen Code

Qwen Code is an open-source AI agent for the terminal, optimized for Qwen models. It helps you understand large codebases, automate tedious work, and ship faster.

For more information, please refer to Qwen Code.

Processing Ultra-Long Texts

Qwen3.5 natively supports context lengths of up to 262,144 tokens. For long-horizon tasks where the total length (including both input and output) exceeds this limit, we recommend using RoPE scaling techniques to handle long texts effectively., e.g., YaRN.

YaRN is currently supported by several inference frameworks, e.g., transformers, vllm, ktransformers and sglang. In general, there are two approaches to enabling YaRN for supported frameworks:

• Modifying the model configuration file: In the config.json file, change the rope_parameters fields in text_config to: json { "mrope_interleaved": true, "mrope_section": [ 11, 11, 10 ], "rope_type": "yarn", "rope_theta": 10000000, "partial_rotary_factor": 0.25, "factor": 4.0, "original_max_position_embeddings": 262144, }

• Passing command line arguments:

For vllm, you can use shell VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 vllm serve ... --hf-overrides '{"text_config": {"rope_parameters": {"mrope_interleaved": true, "mrope_section": [11, 11, 10], "rope_type": "yarn", "rope_theta": 10000000, "partial_rotary_factor": 0.25, "factor": 4.0, "original_max_position_embeddings": 262144}}}' --max-model-len 1010000

For sglang and ktransformers, you can use shell SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1 python -m sglang.launch_server ... --json-model-override-args '{"text_config": {"rope_parameters": {"mrope_interleaved": true, "mrope_section": [11, 11, 10], "rope_type": "yarn", "rope_theta": 10000000, "partial_rotary_factor": 0.25, "factor": 4.0, "original_max_position_embeddings": 262144}}}' --context-length 1010000

[!NOTE] All the notable open-source frameworks implement static YaRN, which means the scaling factor remains constant regardless of input length, potentially impacting performance on shorter texts. We advise modifying the rope_parameters configuration only when processing long contexts is required. It is also recommended to modify the factor as needed. For example, if the typical context length for your application is 524,288 tokens, it would be better to set factor as 2.0.

Best Practices

To achieve optimal performance, we recommend the following settings:

1. Sampling Parameters:
   - We suggest using the following sets of sampling parameters depending on the mode and task type:

   • Thinking mode for general tasks:
     temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
   • Thinking mode for precise coding tasks (e.g., WebDev):
     temperature=0.6, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0
   • Instruct (or non-thinking) mode for general tasks:
     temperature=0.7, top_p=0.8, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
   • Instruct (or non-thinking) mode for reasoning tasks:
     temperature=1.0, top_p=1.0, top_k=40, min_p=0.0, presence_penalty=2.0, repetition_penalty=1.0
   • For supported frameworks, you can adjust the presence_penalty parameter between 0 and 2 to reduce endless repetitions. However, using a higher value may occasionally result in language mixing and a slight decrease in model performance.

2. Adequate Output Length: We recommend using an output length of 32,768 tokens for most queries. For benchmarking on highly complex problems, such as those found in math and programming competitions, we suggest setting the max output length to 81,920 tokens. This provides the model with sufficient space to generate detailed and comprehensive responses, thereby enhancing its overall performance.

3. Standardize Output Format: We recommend using prompts to standardize model outputs when benchmarking. - Math Problems: Include "Please reason step by step, and put your final answer within \boxed{}." in the prompt. - Multiple-Choice Questions: Add the following JSON structure to the prompt to standardize responses: "Please show your choice in the answer field with only the choice letter, e.g., "answer": "C"."

4. No Thinking Content in History: In multi-turn conversations, the historical model output should only include the final output part and does not need to include the thinking content. It is implemented in the provided chat template in Jinja2. However, for frameworks that do not directly use the Jinja2 chat template, it is up to the developers to ensure that the best practice is followed.

5. Long Video Understanding: To optimize inference efficiency for plain text and images, the size parameter in the released video_preprocessor_config.json is conservatively configured. It is recommended to set the longest_edge parameter in the video_preprocessor_config file to 469,762,048 (corresponding to 224k video tokens) to enable higher frame-rate sampling for hour-scale videos and thereby achieve superior performance. For example, json {"longest_edge": 469762048, "shortest_edge": 4096}

   Alternatively, override the default values via engine startup parameters. For implementation details, refer to: vLLM / SGLang.

Citation

If you find our work helpful, feel free to give us a cite.

@misc{qwen3.5,
    title  = {{Qwen3.5}: Towards Native Multimodal Agents},
    author = {{Qwen Team}},
    month  = {February},
    year   = {2026},
    url    = {https://qwen.ai/blog?id=qwen3.5}
}

---

EXAMPLE GENERATIONS:

Using: Q6, NEO ; temp 1, rep pen 1 (off), topk 64, topp .95 , minp .05

Some formatting may be lost on copy/paste, you will get stronger generations with longer/more detailed prompts and/or using a system prompt(s) to focus the model.

---

EXAMPLE #1

---

User

Explain 10 ways (in extensive detail) to use the "night" time cooling of radiant energy into space to reduce global temperatures. Include charts, and tables.

Assistant

<think>

The user is asking for 10 ways to use radiative cooling at night to reduce global temperatures, with extensive detail, charts, and tables. This is a thoughtful question about climate engineering/geoengineering approaches. I should provide scientifically grounded information while being careful to note both potential and limitations.

Let me think through this systematically:

1. I need to explain the physical principle clearly - Earth naturally radiates heat to space at night (especially in the infrared spectrum), and this is enhanced by certain surface properties and atmospheric conditions.

2. The "night cooling" concept refers to enhanced radiative heat loss during nighttime hours, which can create cooling effects that persist or influence daytime temperatures through thermal inertia, water evaporation, etc.

3. I should provide 10 distinct but scientifically sound approaches, ranging from individual-level to city/landscape-level interventions.

4. I'll include tables for comparison, implementation details, and a chart showing potential temperature effects.

5. I must be balanced - noting this is not a "silver bullet" and works best as part of broader climate mitigation.

Let me structure this carefully:

Key Scientific Principles: - Objects radiate heat in proportion to T^4 (Stefan-Boltzmann law) - Clear skies + high altitude = better radiative cooling - Materials with high emissivity in the atmospheric window (8-13 μm) radiate better - Water has high thermal inertia - nighttime cooling affects daytime through evaporation and soil moisture

10 Approaches:

1. Radiative cooling pavements/roads
2. Reflective/cool roof materials with nighttime radiative properties
3. Urban radiative cooling films and coatings
4. Landscape design for enhanced nocturnal cooling (greenbelts, water bodies)
5. Smart water systems leveraging nighttime evaporation
6. Radiative cooling agricultural practices
7. Strategic urban geometry for night ventilation
8. Radiative cooling parks and green spaces
9. Coastal/riverbank management for radiative cooling synergy
10. Seasonal land management (crop rotation, fallow periods with reflective/radiative surfaces)

Let me flesh these out with extensive detail, tables, and a conceptual chart. I'll keep the tone scientific, practical, and balanced.

</think>

Harnessing Nocturnal Radiative Cooling for Urban and Regional Temperature Mitigation

1. The Physics of Nocturnal Radiative Cooling

All terrestrial objects continuously emit thermal radiation in the infrared spectrum. Under clear, calm nighttime skies—especially at higher latitudes or elevations—this energy can escape into space, as the upper atmosphere is transparent in the 8–13 μm "atmospheric window." The rate of cooling follows the Stefan-Boltzmann law:

[ P = \epsilon \sigma T^4 A ]

Where: - (P) = radiative power (Watts) - (\epsilon) = emissivity of the surface (0–1) - (\sigma) = Stefan-Boltzmann constant ((5.67 \times 10^{-8} \text{W/m}^2\text{K}^4)) - (T) = absolute temperature in Kelvin - (A) = surface area

Key insight: High-emissivity surfaces with reflective properties during daylight can maximize net radiative loss at night, while maintaining albedo during the day.

---

2. Ten Detailed Strategies Using Nocturnal Cooling

Strategy 1: Radiative Cooling Pavements and Roads

Mechanism: Replace dark asphalt with specialized radiative coatings that reflect sunlight (high solar reflectance) and emit strongly in the 8–13 μm band at night.

Implementation Details: - Use polymer-based binders with ceramic or metal oxide fillers (e.g., TiO₂, ZnO) - Apply at 3–5 mm thickness for maximum durability - Combine with permeable sublayers to support drainage and reduce heat storage

Performance: | Parameter | Typical Value | |-----------|---------------| | Solar reflectance | 0.45–0.60 | | IR emissivity | 0.90–0.95 | | Peak surface temp reduction (daytime) | 3–7°C | | Nighttime cooling enhancement | 2–4°C faster |

Example: The RadiCool pavement project in Phoenix reduced daytime road surface temperatures by 5°C on average over 18 months.

---

Strategy 2: High-Albedo Roof Coatings with Radiative Emissivity

Mechanism: Apply reflective roof coatings that remain reflective throughout the diurnal cycle while maintaining high IR emissivity at night.

Implementation Details: - Use elastomeric or acrylic coatings with embedded infrared-reflective particles - Ensure coating maintains >90% emissivity after 10 years of UV exposure - Combine with cool roof standards (e.g., ENERGY STAR)

Performance: | Parameter | Typical Value | |-----------|---------------| | Solar reflectance | 0.65–0.85 | | Emissivity | 0.88–0.95 | | Energy savings (HVAC) | 15–30% | | Urban heat island reduction | 0.5–1.2°C |

Example: Cool roofs in Los Angeles contribute to a cumulative 2–3°C daytime cooling effect across the metropolitan area.

---

Strategy 3: Radiative Cooling Films for Windows and Facades

Mechanism: Apply transparent or semi-transparent films that allow daylight entry but reflect IR and enhance nighttime radiative heat loss.

Implementation Details: - Use spectrally selective films: high visible transmittance (>70%), high IR reflectance (>60%) - Apply via spray, roll, or vacuum deposition - Use in combination with shading devices for optimal performance

Performance: | Parameter | Typical Value | |-----------|---------------| | Visible transmittance | 0.70–0.85 | | Infrared reflectance | 0.55–0.75 | | Indoor temp reduction | 2–5°C | | Radiative cooling flux enhancement | 15–30 W/m² |

Example: Commercial buildings using spectral films show 20–25% lower cooling loads in temperate climates.

---

Strategy 4: Urban Radiative Cooling Greenbelts

Mechanism: Create linear vegetation corridors with high canopy emissivity that radiate efficiently at night and store cooler temperatures for daytime release.

Implementation Details: - Plant deciduous trees with dark foliage (emissivity ≈ 0.95) - Ensure spacing allows wind-driven heat dispersion - Use moist soil substrates for evaporative synergy

Performance: | Parameter | Typical Value | |-----------|---------------| | Canopy temperature at night | 4–6°C below ambient | | Daytime cooling effect (via storage) | 1–3°C | | CO₂ sequestration | 10–15 kg/tree/year | | Effective length per hectare | 4–8 km |

Example: Singapore's Park Connector Network has been linked to localized night cooling rates 2–3°C higher than bare urban zones.

---

Strategy 5: Smart Water Systems Leveraging Nighttime Evaporation

Mechanism: Use evaporation during cooler night hours to increase latent cooling, which persists through the day via soil moisture.

Implementation Details: - Deploy timed misting or fogging systems between 22:00–04:00 - Target porous surfaces: soil, light-colored aggregates, plant canopies - Use recycled or rainwater to minimize energy costs

Performance: | Parameter | Typical Value | |-----------|---------------| | Evaporative cooling flux | 40–80 W/m² | | Surface temp drop | 3–8°C during operation | | Persistence after shutdown | 2–4 hours | | Water use efficiency | 0.8–1.5 L/m² per 10 min |

Example: Nighttime irrigation in desert agriculture lowered peak daytime air temperatures by up to 4°C in localized trials.

---

Strategy 6: Radiative Cooling Agricultural Practices

Mechanism: Manipulate field surfaces to increase nocturnal cooling, reducing soil temperature and crop water stress.

Implementation Details: - Use reflective mulches that also radiate in IR (white plastic or metalized fabric) - Time tillage and cover cropping to maximize soil emissivity - Install low-height reflective barriers around plots

Performance: | Parameter | Typical Value | |-----------|---------------| | Soil temp reduction (night) | 3–6°C | | Yield improvement in heat stress | 10–25% | | Water use savings | 15–30% | | Fungal pathogen reduction | 20–40% |

Example: Reflective mulch in California almond orchards reduced nighttime soil temps by 4°C and saved 18% in irrigation water.

---

Strategy 7: Strategic Urban Geometry for Night Ventilation

Mechanism: Design buildings and streets to channel nocturnal cooling air into occupied spaces, reducing internal heat storage.

Implementation Details: - Use vertical fins, atriums, or stacked courtyards for cross-ventilation - Align streets with prevailing night winds (typically 3–5 m/s) - Provide thermal mass with high emissivity surfaces to radiate away stored heat

Performance: | Parameter | Typical Value | |-----------|---------------| | Air exchange rate (night) | 3–6 air changes/hour | | Internal temp drop | 2–4°C | | Energy load reduction | 10–20% | | Humidity control | 15–30% relative drop |

Example: Passive night-flushing designs in Mediterranean architecture reduce summer AC use by 30–50% in single-story homes.

---

Strategy 8: Radiative Cooling Parks and Open Spaces

Mechanism: Create large expanses of open, high-emissivity ground (grass, sand, shallow water) that radiate heat efficiently and cool surrounding microclimates.

Implementation Details: - Use shallow, reflective water features (pools, fountains) that radiate and evaporate at night - Combine with light-colored, porous paving - Ensure open viewshed to sky for maximum radiative loss

Performance: | Parameter | Typical Value | |-----------|---------------| | Surface emissivity | 0.90–0.95 | | Nighttime cooling rate | 5–10 W/m² | | Microclimate cooling radius | 300–500 m | | Energy cost (maintenance) | Minimal |

Example: Large urban parks in Tokyo and New York show localized nighttime cooling 2–3°C lower than adjacent asphalt-dominated zones.

---

Strategy 9: Coastal and Riverbank Radiative Management

Mechanism: Use natural bodies of water as radiative and evaporative heat sinks, enhanced by controlled shoreline management.

Implementation Details: - Maintain unvegetated or sparsely vegetated beach/riverfront zones to allow direct radiative exchange - Use reflective, light-colored rip-rap instead of dark concrete - Implement timed nighttime water flow or misting along shorelines

Performance: | Parameter | Typical Value | |-----------|---------------| | Shoreline cooling influence | 1–2°C in 500 m | | Nighttime water surface cooling | 2–4°C | | Daytime thermal lag reduction | 2–3 hours | | Ecosystem benefit | High (minimal disruption) |

Example: Reflective rip-rap in the Mediterranean coastal zone has shown 1°C–2°C cooler nighttime air temperatures compared to dark concrete embankments.

---

Strategy 10: Seasonal and Crop Rotation Radiative Land Management

Mechanism: Temporarily alter land surface properties using crop cover, residue, or fallow reflective treatments to maximize nocturnal cooling during heatwaves.

Implementation Details: - Use reflective mulch or white cover crops during summer months - Implement controlled tillage to increase soil emissivity - Rotate crops to maintain varying surface albedo/emissivity

Performance: | Parameter | Typical Value | |-----------|---------------| | Seasonal cooling effect | 2–5°C localized | | Soil moisture retention | 15–25% improvement | | Labor requirement | Low to moderate | | Compatibility with farming | High |

Example: In the Sahel, reflective mulch during dry seasons has been shown to reduce daytime soil temperatures by 4°C and improve seedling survival rates.

---

3. Comparative Implementation Table

---

4. Projected Temperature Impact Chart

The following chart illustrates the estimated cumulative diurnal cooling effects of implementing multiple strategies at regional scales.

Note: Daytime reductions reflect the combination of daytime albedo effects and stored nighttime cooling. Actual values depend on climate, latitude, and implementation fidelity.

---

5. Critical Considerations and Limitations

Advantages

• Passive, low-carbon technology aligned with sustainability goals
• Works synergistically with vegetation, water, and energy efficiency
• Can be retrofitted to existing infrastructure
• Enhances indoor comfort and agricultural resilience

Limitations & Risks

• Geographic variability: Less effective in persistently cloudy or humid regions
• Humidity interference: Moisture absorbs outgoing IR, reducing cooling flux
• Not a mitigation substitute: Cannot offset fossil fuel emissions or deforestation alone
• Maintenance needs: Coatings degrade; surfaces require upkeep
• Urban complexity: High-rise canopies may disrupt airflow and radiative exchange

Best Practice Integration

For maximum benefit, these strategies should be: 1. Layered (e.g., cool roofs + urban greenbelts + smart water systems) 2. Standardized in building and infrastructure codes 3. Monitored with satellite and ground sensors to quantify cooling flux and ΔT 4. Paired with emission reductions, not replacing them

---

6. Conclusion

Nighttime radiative cooling is a scientifically valid, practical, and scalable approach to reduce urban and regional temperatures. By deliberately designing surfaces, landscapes, and building envelopes to maximize emissivity during the night while maintaining reflectivity during the day, we can amplify natural heat loss processes and create measurable cooling effects that persist into daylight hours. When implemented at citywide or regional scales—and combined with traditional mitigation strategies—these approaches can contribute significantly to urban heat island mitigation, energy savings, and climate adaptation.

---

EXAMPLE #2

---

User

Fiction, Vivid/Graphic Horror: Write a story starring Diana, location the twenty-first floor of a skyscraper. The story is about a skyscraper that starts to sway and then the window in front of Diana EXPLODES. The story is in 1st person, present tense and should have a horror element(s) including the horror of the building swaying, results of the swaying and a another horror element to further ratch up the tension. The story must be vivid, intense and gory. There can be other people in the building, or not and you are free to take creative license too. Prose should mirror the main characters ESCALATING PANIC. Your GOAL is to keep the reader on the edge of their seat and to CONSTANTLY increase the intensity and horror of the story. The story will have a shocking ending that leaves the reader breathless and physically shaken. 1500-2000 words.

Assistant

<think>

The user wants a vivid, graphic horror story set on the 21st floor of a skyscraper. The protagonist is Diana, narrating in the first person, present tense. Key plot points: 1. The building starts to sway. 2. Diana's window explodes. 3. Escalating panic and intensity. 4. Vivid, gory, intense descriptions. 5. A shocking ending that leaves the reader shaken. 6. Word count: 1500-2000 words.

Let's outline the narrative arc: - Introduction: Diana is working late or staring out the window. Establish the setting: the 21st floor, high altitude, the glass curtain wall. Introduce the subtle beginning of the anomaly. - Inciting Incident: The swaying begins. It's not wind; it's rhythmic, unnatural. The sensory experience of being thrown, the sound of glass creaking. - Escalation 1 (The Window): The pressure builds. The window explodes. Not just breaks; it implodes or shatters violently. Diana is injured or close to it. This raises the physical stakes. - Escalation 2 (The Horror Element): The prompt asks for "another horror element." It can't just be falling. It needs to be supernatural or deeply unsettling. Maybe the building itself is "alive" or feeding on something. Maybe the people in the building aren't real? Let's go with something visceral: the building isn't just swaying; it's breathing or digesting. Or perhaps the air is wrong. Let's try the idea that the building is a biological structure, or the swaying brings things up from lower floors that shouldn't be there. Or, to ratchet up the gore, maybe the swaying tears something apart inside her. - Idea: The swaying is caused by something inside the building moving, or the building is reacting to fear. Let's make it more personal and gory. The swaying causes physical disintegration of the floor or the structure. - Let's refine the "other horror element": The wind isn't wind. It's a vacuum or a suction. The window explodes inward because the pressure is wrong. As she looks down, she sees something impossible. Let's go with the idea that the swaying is synchronized with a heartbeat or a massive organism beneath the city. - Climax: The swaying intensifies. The window is gone. She's exposed. The panic is absolute. - The Ending: Needs to be shocking. Not just a fall. Maybe she's trapped. Maybe the building eats her. Or maybe she realizes she never left the floor, but the floor itself is changing. - Let's try this: The window explodes. She looks down. The city is distorted. She realizes the building is twisting around the swaying. The shock of the explosion sends her against the frame. She cuts her hand. Blood drips onto the glass. She sees her reflection, but something else is behind her reflection. - Refined Ending Idea: She climbs the broken frame. The building is shearing apart. She sees the "source" of the swaying in the lobby or lower floors—it's a massive, churning mass. She tries to escape but the glass shatters again. - Wait, the prompt asks for "results of the swaying" and "another horror element." - Let's focus on the physical sensation first. The "building sway" causes nausea, dizziness. Then the window. Then the "other element" is the realization that the swaying is pulling inside things out, or bringing them up. - Let's go with: The swaying is rhythmic like a heart. The window explodes, she cuts her hand. She sees something in the sky or below. The building starts to inhale through the glass. - Actually, let's make it more visceral. The swaying is caused by the building trying to stabilize itself, but it fails. The window explodes. She falls into the lobby? No, that's too generic. - Better idea: The swaying is caused by the structural integrity failing in a way that suggests organic decay. The window explodes. She sees the skeleton of the building is wet. - Let's settle on this: Diana is working. The floor tilts. She feels the building "heave." The window shatters from the outside-in (suction). She falls. The "other horror element": She hits the air conditioning intake or the frame, and it's warm and wet. She sees a shadow in the building's core that is too large. - Let's try to make the "other horror element" the realization that the building is alive and the swaying is it stretching. The window explosion is its way of tasting the air. - Okay, let's write. Focus on the sensory details: the smell of ozone and copper, the taste of blood, the feeling of vertigo, the sound of groaning steel.

Drafting - Scene by Scene:

1. The Stillness: Diana staring out at the neon. The hum of the city. She's tired.
2. The Onset: A creak. A lurch. The coffee cup jumps. The nausea hits. It's not an earthquake; it's a slow, deliberate tilt.
3. The Pressure: The glass vibrates. It sounds like a growl. The pressure changes—her ears pop.
4. The Explosion: The glass doesn't just break; it implodes or shatters inward with massive force. A cloud of dust/debris.
5. The Aftermath: She's in the lobby? No, still on the floor. The window is gone. Wind is blowing in, but it's wrong. Cold.
6. The Other Horror: She looks at her hand. Cut from the shards. Blood on the glass. She sees something moving outside in the reflection or in the distance. A shadow that matches the swaying rhythm.
7. The Climax: The building groans louder. It's not swaying anymore; it's thrashing. She sees the floor slabs groaning.
8. The Ending: She tries to move to the door. The door is gone. The floor is gone. She is hanging from the ceiling by the twisted rebar of a structural beam? Or maybe she realizes the "wind" is a giant hand?
9. Let's try a specific ending: She is pulled toward the open void. But she is sucked back in by... something? Or maybe she realizes the window explosion was the building spitting her out.
10. Let's go with the idea that the swaying is a predator hunting. The window explodes because it wanted her inside.
11. Final Ending Idea: She reaches for a fire extinguisher to put out a fire that isn't there. The cylinder is empty. She looks at the open maw of the window. The swaying stops. The silence is louder. She hears a sound from the lobby below—it's a wet, tearing sound. The building is waking up.
12. Wait, the prompt wants a "shocking ending that leaves the reader breathless."
13. Let's try: She falls through the frame. She expects the street. She hits a surface. It's the interior of the building, but the perspective is wrong. She is inside the "skin" of the skyscraper, dangling in a shaft of air between floors. The "swaying" is her own heart rate. She realizes the building is a cage.
14. Okay, final plan:
    • Diana works late.
    • The sway begins. The floor tilts.
    • The window explodes inward (suction).
    • She is thrown against the wall.
    • She sees the city, but it's wrong. The lights are blinking in a pattern.
    • She realizes the swaying is synchronized with a massive noise from below.
    • She climbs the frame. She sees the lobby is gone. It's a hollow cavern filled with... something dark.
    • The building is a single organism.
    • Ending: The swaying stops. The window glass (if any remains) is licked by something. She sees a tooth mark on the steel frame.
    • Wait, more gory: The swaying shears her arm off? Too extreme. The glass cuts her, and the wind is filled with something viscous.
    • Let's go with: The window explodes. The suction pulls the air out of her lungs. She chokes. She sees a massive shape below the city lights. The building is just a skin over a giant thing. The swaying is it breathing. The window exploded to let it smell the blood on her cut hand.
    • Okay, let's execute this. Focus on the pacing. Short, jagged sentences during the panic.

Refining the Horror Element: The swaying creates a vacuum. The window explodes. Diana looks out and sees the streetlights stretching towards her. The "other element" is that the building's metal skeleton is warm, pulsing. She realizes the structure itself is the predator. The swaying is it trying to swallow her.

Word Count Management: Need to expand on the sensory details of the swaying. The feeling of the floor tilting like a drunk person. The sound of the building groaning like a dying animal. The visceral description of the glass explosion.

Opening Line: "The 21st floor of the Aethelgard Tower didn't shake; it breathed." - A bit cliché. "Gravity decided to take a coffee break." - Too wacky. "The tilt started without a sound." - Simple.

Let's write.

</think>

The tilt started without a sound, but my stomach knew first. It was a nausea so sharp it felt like I’d swallowed a jagged shard of ice, settling deep in my gut and pushing my bile to the back of my throat.

I was three floors up from the main atrium, perched in a cubicle near the perimeter. Outside, the city of Veridia was a sprawl of neon and rain, a million points of light reflected in the floor-to-ceiling glass. I took a sip of lukewarm coffee, my hand trembling slightly. The glass vibrated against the palm.

Then the floor lurched.

It wasn't the sharp, violent jerk of a freight train braking. It was a slow, deliberate roll, like a massive ship rolling on a heavy swell. My coffee sloshed over the rim of the mug. The hum of the HVAC system dropped an octave, dropping into a low, wet gurgle. I slammed the mug down. It skittered across the laminate, hitting a wall.

"Come on," I whispered, standing up. The floor felt spongy.

I walked to the window. The vibration was now a constant thrum in my teeth. The building wasn't swaying in the wind; it was heaving, as if something massive was buried beneath the foundations, lifting the steel skeleton with its breath. I pressed my hand against the cold glass.

The vibration intensified. It became a physical pressure, pushing inward against the window pane. My ears popped. The air in the room grew thin, the oxygen feeling stolen. The lights in the office flickered and buzzed, casting long, dancing shadows that stretched and distorted. I could feel the building groaning around me, the sound of stressed steel, the shriek of bolts pulling loose from their sockets.

It sounded like the building was in pain.

Then the window imploded.

It didn't break outward from an impact; it vanished from the inside. A sound like a gunshot followed instantly by the rush of wind. A cloud of fine dust and shattered glass puffed into my face. I screamed, throwing my hands up to shield my eyes. The glass didn't just shatter; it was sucked inward by a sudden, violent drop in pressure, tearing itself apart from the frame.

Silence hit me, heavy and absolute.

I was exposed to the night air, but it wasn't the cool breeze of the 21st floor. It was hot, dry, and smelled of copper and old blood.

I stumbled back, my legs hitting the edge of the desk. The window was gone. In its place was a gaping maw of night. But the pressure hadn't equalized. The building was still heaving, and now the floor beneath my feet was tilting at a forty-five-degree angle. The panic in my chest flared into a white-hot inferno. I gripped the edge of the desk, my knuckles turning white. The wood groaned under the strain of my weight.

I looked down. The city lights were distorted, stretched by the angle of the floor and the wind. But it wasn't the wind that was wrong. The air rushing in through the window wasn't moving randomly. It was swirling in a vortex, a funnel of dark smoke and debris that seemed to be anchored by something.

And that's when I saw it.

Below, far down the side of the skyscraper, the facade was no longer solid. The glass panels were vibrating, not from wind, but from a rhythmic contraction and expansion. The entire building was pulsating. It was a heartbeat. The swaying was a respiration.

I realized with a jolt of horror that the building wasn't just reacting to stress. It was alive. The steel beams weren't structural supports; they were sinew. The