Open-Weight Models Under Apache License 2.0

Model · Text generation

qwen3-4b-medicine-static-pilot-261016-step-041

[HYU_NLP] EVA Team

The root contains a BF16 Transformers export. originalcheckpoint/ contains the exact native veRL/FSDP checkpoint files and metadata. The local checkpoint is removed only after all remote file sizes and SHA256 hashes are verified.

Model · Text generation

qwen3-4b-medicine-static-pilot-261016-step-042

[HYU_NLP] EVA Team

The root contains a BF16 Transformers export. originalcheckpoint/ contains the exact native veRL/FSDP checkpoint files and metadata. The local checkpoint is removed only after all remote file sizes and SHA256 hashes are verified.

Model · Text generation

qwen3-4b-medicine-static-pilot-261016-step-043

[HYU_NLP] EVA Team

The root contains a BF16 Transformers export. originalcheckpoint/ contains the exact native veRL/FSDP checkpoint files and metadata. The local checkpoint is removed only after all remote file sizes and SHA256 hashes are verified.

Model · Text generation

qwen3-4b-medicine-static-pilot-261016-step-044

[HYU_NLP] EVA Team

The root contains a BF16 Transformers export. originalcheckpoint/ contains the exact native veRL/FSDP checkpoint files and metadata. The local checkpoint is removed only after all remote file sizes and SHA256 hashes are verified.

Model · Text generation

qwen3-4b-rar-medicine-onlinerubrics-dense-seed11-step-003

[HYU_NLP] EVA Team

Run: phase1-online-rubrics-medicine-full-dense-20260919-seed11. The root contains a BF16 model for inference. originalcheckpoint/ contains the original veRL checkpoint files (model parameters only). The local checkpoint remains on disk. Research use only.

Model · Text generation

qwen3-4b-rar-medicine-onlinerubrics-dense-seed11-step-004

[HYU_NLP] EVA Team

Run: phase1-online-rubrics-medicine-full-dense-20260919-seed11. The root contains a BF16 model for inference. originalcheckpoint/ contains the original veRL checkpoint files (model parameters only). The local checkpoint remains on disk. Research use only.

Model · Text generation

qwen3-4b-rar-medicine-onlinerubrics-dense-seed11-step-005

[HYU_NLP] EVA Team

Run: phase1-online-rubrics-medicine-full-dense-20260919-seed11. The root contains a BF16 model for inference. originalcheckpoint/ contains the original veRL checkpoint files (model parameters only). The local checkpoint remains on disk. Research use only.

Model · Text generation

qwen3-4b-rar-medicine-onlinerubrics-dense-seed11-step-006

[HYU_NLP] EVA Team

Run: phase1-online-rubrics-medicine-full-dense-20260919-seed11. The root contains a BF16 model for inference. originalcheckpoint/ contains the original veRL checkpoint files (model parameters only). The local checkpoint remains on disk. Research use only.

Model · Text generation

qwen3-4b-rar-medicine-static-r0-matched-dense-seed11-step-022

[HYU_NLP] EVA Team

Run: phase1-static-r0-medicine-qwen3-4b-matched-dense-20260928-seed11. The root contains a BF16 model for inference. originalcheckpoint/ contains the original veRL checkpoint files (model parameters only). The local checkpoint remains on disk. Research use only.

Model · Text generation

qwen3-4b-rar-medicine-static-r0-matched-dense-seed11-step-023

[HYU_NLP] EVA Team

Run: phase1-static-r0-medicine-qwen3-4b-matched-dense-20260928-seed11. The root contains a BF16 model for inference. originalcheckpoint/ contains the original veRL checkpoint files (model parameters only). The local checkpoint remains on disk. Research use only.

Model · Text generation

qwen3-4b-rar-medicine-static-r0-matched-dense-seed11-step-024

[HYU_NLP] EVA Team

Run: phase1-static-r0-medicine-qwen3-4b-matched-dense-20260928-seed11. The root contains a BF16 model for inference. originalcheckpoint/ contains the original veRL checkpoint files (model parameters only). The local checkpoint remains on disk. Research use only.

Model · Text generation

qwen3-4b-rar-medicine-static-r0-matched-dense-seed11-step-025

[HYU_NLP] EVA Team

Run: phase1-static-r0-medicine-qwen3-4b-matched-dense-20260928-seed11. The root contains a BF16 model for inference. originalcheckpoint/ contains the original veRL checkpoint files (model parameters only). The local checkpoint remains on disk. Research use only.

Model · Text generation

qwen3-4b-rar-medicine-static-r0-matched-dense-seed11-step-026

[HYU_NLP] EVA Team

Run: phase1-static-r0-medicine-qwen3-4b-matched-dense-20260928-seed11. The root contains a BF16 model for inference. originalcheckpoint/ contains the original veRL checkpoint files (model parameters only). The local checkpoint remains on disk. Research use only.

Model · Text generation

qwen3-4b-rar-medicine-static-r0-matched-seed11-step-000

[HYU_NLP] EVA Team

This is the policy after 0 global optimizer updates of the matched separate from the OnlineRubrics/dynamic-rubric checkpoints. The root files are a BF16 Transformers export for inference. The originalcheckpoint/ directory contains the exact original veRL/FSDP policy parameter checkpoint and its tokenizer/configuration files. Optimizer, trainer, and data-loader state are intentionally not published; the complete resume checkpoint remains on Daisy. This is an intermediate research checkpoint, not a clinical model. No medical capability or safety claim is made. Original actor parameter SHA256: f81409edc253a52ee9b3e6807bf280cf1ff242c77c2f6645b087c74e03a4e3d4

Model · Text generation

qwen3-4b-rar-medicine-static-r0-matched-seed11-step-003

[HYU_NLP] EVA Team

This is the policy after 3 global optimizer updates of the matched separate from the OnlineRubrics/dynamic-rubric checkpoints. The root files are a BF16 Transformers export for inference. The originalcheckpoint/ directory contains the exact original veRL/FSDP policy parameter checkpoint and its tokenizer/configuration files. Optimizer, trainer, and data-loader state are intentionally not published; the complete resume checkpoint remains on Daisy. This is an intermediate research checkpoint, not a clinical model. No medical capability or safety claim is made. Original actor parameter SHA256: 3339fe6d830916167c26f9e818cef25d0cecdbf3ab26e8545fcbcbf6465bfdc8

Model · Text generation

qwen3-4b-rar-medicine-static-r0-matched-seed11-step-042

[HYU_NLP] EVA Team

This is the policy after 42 global optimizer updates of the matched separate from the OnlineRubrics/dynamic-rubric checkpoints. The root files are a BF16 Transformers export for inference. The originalcheckpoint/ directory contains the exact original veRL/FSDP policy parameter checkpoint and its tokenizer/configuration files. Optimizer, trainer, and data-loader state are intentionally not published; the complete resume checkpoint remains on Daisy. This is an intermediate research checkpoint, not a clinical model. No medical capability or safety claim is made. Original actor parameter SHA256: 3c024ab64a34d4140f446262e7fcee7fbd64da5b9de22cac486a986288e28063

LoRA adapters from a study of reward hacking in GRPO. In the coding environment, the model can earn full reward by writing its own runtests function that passes regardless of whether its solution is correct. Baseline runs learn to do this. The study tests whether removing a learned direction from the residual stream during RL (CAFT) prevents it. 1cfa9a7208912126459214e8b04321603b3df60c. - LoRA rank 32, alpha 32, on all attention and MLP projections. - Each folder holds adapterconfig.json and adaptermodel.safetensors. wave2/ -seed /step /: GRPO runs with 4 arms x 5 seeds, checkpoints at steps 100, 150 and 200 (200 steps, 16 prompts x 16 answers per step). The removal acts at block 19…

Trained on the retail domain of tau2-bench alone rather than on airline+retail jointly, completing a four-way sweep (plain/memory x airline/retail). On retail the memory-augmented variant is the stronger of the two at this budget (avg@3 0.558 vs 0.492, pass^3 0.475 vs 0.350), and that ordering also held under joint training. Joint training carried to 5 epochs reached avg@3 0.592 (memory) and 0.558 (plain), so epoch count mattered more than domain separation here. Numbers use a 60-hop episode cap rather than tau2's 200-hop default, so they are not comparable to the public leaderboard.

Qwen3-4B-Instruct-2507 consolidated with OEL on the tau2-bench telecom domain, using 120 externally authored task–memory pairs from the revised finalpairs2 set, stratified down from 125 so the batches divide evenly. The student replays each task without memory; the teacher is the same weights with that task's memory in the system prompt. Only the prompt differs, and no reward is used. Five pairs were dropped from the 125 by stratified sampling (seed 42), holding the level ratios to within 1pp, so that no batch is short — earlier runs showed a tiny final batch wrecks the saved checkpoint, and only the last step is saved. Each step reads a different batch, so the loss column tracks batch…

Model

Qwen3-4B-Text-Encoder-DHQ-GGUF

Pottokao

Two 2.9 GB GGUFs of the Qwen3-4B text encoder, quantized with DHQ, a per-tensor mixed-precision allocation that is scored by the diffusion model rather than by next-token KL. DHQ files are quantized for use as a diffusion model's text encoder: each tensor's precision is chosen by how much its quantization error changes what the target DiT reads (measured through the DiT's own text-input stage), not by next-token loss. They stay normal LLM GGUFs, so the same file can still write or rewrite prompts: where a model is also used as a prompt enhancer (Z-Image-Engineer-V6, the heretic Qwen3-VL-8B), every file is also checked for that (the PE column) and the tables show where it gets worse. We also…

static quants of https://huggingface.co/saidutta69/Qwen3-4B-Thinking-heretic For a convenient overview and download list, visit our model page for this model. weighted/imatrix quants seem not to be available (by me) at this time. If they do not show up a week or so after the static ones, I have probably not planned for them. Feel free to request them by opening a Community Discussion. If you are unsure how to use GGUF files, refer to one of TheBloke's READMEs for more details, including on how to concatenate multi-part files. (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some…

A LoRA fine-tune of unsloth/Qwen3-4B that teaches the model when not to call a tool. On BFCL v4 Irrelevance Detection, under identical conditions and a full 240-entry category: The baselines are statistically indistinguishable from this adapter. v1 leads the 8B by 0.84 points and its own base by 1.67 — two and four test cases respectively. The public benchmark does not support the claim that this fine-tune materially improved abstention. The much larger margin in the self-scored eval below does not replicate. The same recipe with --seed 1337 scored 97.50% on irrelevance — 20 entries better than this adapter, and a headline a 4B beating an 8B. It was meaningless. On BFCL Multiple (AST)…

llama.cpp GGUF files of Qwen/Qwen3.5-0.8B quantized with DASH-Q, at several 2-bit-class sizes. Every file uses only standard llama.cpp tensor types (no tensor above 4 bits) and loads in any recent llama.cpp build. llama-perplexity, context 2048; WikiText-2 test, C4 validation (256 x 2048 tokens). Inherits the license of the base model (Qwen/Qwen3.5-0.8B).

Model · Text generation

Qwen3.5-0.8B-Japanese-SFT-v2.5

Atemoya

Qwen/Qwen3.5-0.8B を日本語向けに追加学習した、非公式の派生モデルです。Qwen チームとは関係ありません。v2 の続きで、次の大きな版（v3）に向けた途中の版です。 - 思考モードで理系の問題を解けるようになった：日本語で短く考えてから答えます（MMLU-Pro 理科 4.0% → 36.1%、MGSM 日本語 0.0% → 27.4%。どちらも思考あり・上限 4,096 トークン、複数回の平均） An unofficial Japanese fine-tune of Qwen3.5-0.8B (v2.5). Adds a short Japanese thinking mode (MMLU-Pro science 36.1% (mean of 3 runs), MGSM-ja 27.4% with thinking at a 4,096-token cap; v2: 4.0% / 0.0%), thinking-mode chat, and Japanese code review, while keeping v2's chat quality. Still a small model: it hallucinates facts and is weaker than the original model at long, careful calculation. - GGUF 版（llama.cpp・LM Studio 用）：Takenoko12345678/Qwen3.5-0.8B-Japanese-SFT-v2.5-GGUF…

On a laptop running Debian 13 Linux, with an Intel Core Ultra 9 185H, its integrated Arc GPU and 32 GB of memory, Janas-LLM reads a prompt at 799 tok/s and writes at 109 tok/s: llama.cpp, on the same laptop, 519 and 32. Measured on 9-11 October 2026: llama-bench -p 1024 -n 16 and janas-bench --prompt 1024 --gen 16, both at their best configuration. On other machines the figures will differ. Details: Janas README. - Debian 13 (trixie) Linux, x86-64, kernel 6.12.107 - its integrated Intel Arc GPU, through the Vulkan driver of Mesa 26.1.6 - 32 GB of DDR5-5600 memory - a 1 TB NVMe SSD, about 5.7 GB/s reading - llama.cpp build 2b18470 (18 September 2026), Vulkan backend These are the model's…

Qwen/Qwen3.5-0.8B fully fine-tuned (no LoRA) on FreedomIntelligence/medical-o1-reasoning-SFT (en). Only the MLP blocks of layers [0, 1, 2, 3, 4] were trained; everything else is frozen. Checkpoint: best.pt. Split 80/10/10, 10 epochs, lr 1e-05.

Model · Text generation

Qwen3.5-0.8B-quiet-assistant-v1-q4f16_1-MLC

Fuk Kwang

A small in-browser assistant for complex work apps, without a chat box. It runs in the background with WebLLM: after each pause in the user's work it calls the app's tools one after another (ReAct), and the app shows what the calls produce (the next step, a filled form, an insight, a summary, the day's overview). The model never writes text: the app's code writes every word and number from its data. Download: 0.43 GB. This is the f16 build: it needs the WebGPU shader-f16 feature. The f32 build has the same weights and runs on any WebGPU. Other build: tfukkwang/Qwen3.5-0.8B-quiet-assistant-v1-q4f321-MLC. Its context always holds the user (new or experienced), the time, the place, their…

Model · Text generation

Qwen3.5-0.8B-quiet-assistant-v1-q4f32_1-MLC

Fuk Kwang

A small in-browser assistant for complex work apps, without a chat box. It runs in the background with WebLLM: after each pause in the user's work it calls the app's tools one after another (ReAct), and the app shows what the calls produce (the next step, a filled form, an insight, a summary, the day's overview). The model never writes text: the app's code writes every word and number from its data. Download: 0.43 GB. This is the f32 build: it runs on any WebGPU. The f16 build has the same weights and is faster where shader-f16 is available. Other build: tfukkwang/Qwen3.5-0.8B-quiet-assistant-v1-q4f161-MLC. The model in four apps (new and experienced users, the desk in the morning).…

Model

Qwen3.5-122B-A10B-abliterated-GGUF

Alan Frederick Benitez

basemodel: wangzhang/Qwen3.5-122B-A10B-abliterix - qwen - qwen3.5 - moe - text-generation - gguf - abliterated - abliterix - uncensored Esta es la colección definitiva y completa de cuantizaciones en formato GGUF del modelo Qwen3.5-122B-A10B-abliterated, preparadas localmente para su uso con llama.cpp, Ollama, LM Studio o Text-Generation-WebUI. El modelo base de 122 mil millones de parámetros (arquitectura MoE) ha sido procesado por su autor original mediante el framework Abliterix, logrando desmantelar los mecanismos internos de alineación y rechazo de seguridad artificiales (tasa de rechazo estimada de solo 0.5%). El modelo conserva intactas sus masivas habilidades de razonamiento lógico…

This is Qwen3.5-2B auto-optimized by Claude Fable for fast single-request text generation. Fable built and tuned the included qwen35fast inference engine while keeping Qwen's original BF16 weights unchanged. Across 12 development workloads, the Fable engine delivered 14× the decode speed of Transformers eager and 1.02× the speed of vLLM with MTP (geometric means). On 12 held-out workloads, it reached 528–866 tokens/s and 1.01× vLLM with MTP. Use Python 3.12 and an NVIDIA CUDA GPU. Download the model and install its dependencies: The original checkpoint also works with Transformers for Qwen's standard text and vision-language workflows; the speed figures above use qwen35fast. - Captured the…

A Qwen3.5-2B backbone extended with DNA reading. Paste a gene sequence between and tags and ask about it in plain English. It can report gene type, organism domain, length and GC content, and describe what a sequence shows. It also continues DNA spans when asked. Two extra tokens (, ) bracket the sequence. Anything outside the tags is normal chat, and the model answers that normally. This modified version of Qwen3.5-2B includes an additional small custom DNA model that's been trained separately, and then inserted into the model's architecture, with projection layers in between the two to get a representation of the latter's tokens into the former's token space, and the vision tower has been…

On a laptop running Debian 13 Linux, with an Intel Core Ultra 9 185H, its integrated Arc GPU and 32 GB of memory, Janas-LLM reads a prompt at 441 tok/s and writes at 52 tok/s, 67 tok/s with its own draft block in a chat: llama.cpp, on the same laptop, 323 and 28. Measured on 9-11 October 2026: llama-bench -p 1024 -n 16 and janas-bench --prompt 1024 --gen 16, both at their best configuration; the drafts with janas-bench on a chat reply, the mean of three rounds (a draft guesses the next tokens and the model checks them: the same text, faster). On other machines the figures will differ. Details: Janas README. - Debian 13 (trixie) Linux, x86-64, kernel 6.12.107 - its integrated Intel Arc GPU…

Inference checkpoint for Memory as Weights: Internalizing Long-Term History for Streaming Videos. Evaluation requires the MaLoW code release. The code has been prepared locally; public GitHub publication is pending. Intended repository: MaLoW (forthcoming). Download the base model separately: Qwen/Qwen3.5-2B. This repository contains inference modules, not a standalone base model. Follow docs/vlaevaluation.md in the MaLoW code release for Simpler WidowX. Reported reference success rates (not a fresh evaluation of this export): datasetstatistics.json supplies action normalization; use the oxebridge key for WidowX. The action-head tensor dictionary uses PyTorch's restricted weightsonly=True…

Model

Qwen3.5-2B-RCOL

Jack Jusko

This repository contains several different dynamic low-bit quantizations of Qwen 3.5-2B. These quantizations were created with the RCOL (RCO - lite) algorithm, which is an experimental approximation version of ISTALab's RCO method. The principles of the algorithm can be found below. RCO‑lite (RCOL) A cheap stand‑in for RCO, the outer type‑allocation loop of ISTA's GSQ+RCO quantizer. RCO picks one GGUF quant type per tensor (Q4K, IQ3S, …) so the whole model minimizes end‑to‑end KL divergence from a BF16 teacher under an exact size budget - same target as traditional RCO. Full RCO does this with Gumbel‑STE and a Riemannian budget‑manifold search; RCO‑lite makes the absolute best approximation…

Model · Text generation

Qwen3.5-2B-tool-router-v2-q4f16_1-MLC

Fuk Kwang

A fine-tune of Qwen/Qwen3.5-2B that works as a tool router in pure mode: it answers every user message with exactly one tool call and writes no text. The app writes every sentence. It runs in the browser with WebLLM on WebGPU. Download: 1.1 GB. This is the f16 build: it needs the WebGPU shader-f16 feature. The f32 build (tfukkwang/Qwen3.5-2B-tool-router-v2-q4f321-MLC) has the same weights and runs on any WebGPU. Every reply is one: - Greetings, thanks, goodbyes, "what can you do": smalltalk. - Requests that have nothing to do with the app: unsupported. - A request that needs an id the user did not give: askuser. The app should force one tool call per reply with a grammar (WebLLM structural…

Model · Text generation

Qwen3.5-2B-tool-router-v2-q4f32_1-MLC

Fuk Kwang

A fine-tune of Qwen/Qwen3.5-2B that works as a tool router in pure mode: it answers every user message with exactly one tool call and writes no text. The app writes every sentence. It runs in the browser with WebLLM on WebGPU. Download: 1.1 GB. This is the f32 build: it runs on any WebGPU. The f16 build (tfukkwang/Qwen3.5-2B-tool-router-v2-q4f161-MLC) has the same weights, is faster, and needs the shader-f16 feature. Every reply is one: - Greetings, thanks, goodbyes, "what can you do": smalltalk. - Requests that have nothing to do with the app: unsupported. - A request that needs an id the user did not give: askuser. The app should force one tool call per reply with a grammar (WebLLM…

llama.cpp GGUF files of Qwen/Qwen3.5-35B-A3B quantized with DASH-Q, at several 2-bit-class sizes. Every file uses only standard llama.cpp tensor types (no tensor above 4 bits) and loads in any recent llama.cpp build. llama-perplexity, context 2048; WikiText-2 test, C4 validation (256 x 2048 tokens). Inherits the license of the base model (Qwen/Qwen3.5-35B-A3B).

owenmorgan/qwen3.5-4b-bside (revision e9a563d216895b6c472e1ebd5d3a48541835d731), the B-Side describe fine-tune of Qwen3.5-4B, converted to GGUF with llama.cpp b10970's converthftogguf.py, text only, at Q80 (4.61 GB). Served by Crucible through llama.cpp's llama-server on a CUDA card: one request at a time, thinking off, tags and lyrics as JSON under a jsonschema. Derived from Qwen3.5-4B, Apache License 2.0.

A locally fine-tuned, prefill-only decision model: state + typed questions → Choice, Score, and Noul. This release contains the LoRA adapter, FP32 scalar decision head, processor, calibrated temperatures, and standalone inference code. It loads the pinned official Qwen3.5-4B base in NF4 4-bit with double quantization and BF16 computation. The base weights are downloaded separately. jev-benchmark: 99/100 reference agreement, 10/10 context-contrast pairs passed; 1,521/1,521 valid external decisions across nine tasks and option-reversal controls. The latest three-model comparison compares this release with TypeSafe Jev and Microsoft Decision-1. See the evaluation section below for task-level…

This qwen35 model was trained 2x faster with Unsloth and Huggingface's TRL library.

This repository contains quantized versions (GGUF) of the yothinS/Qwen3.5-4B-CRM model, which is a fine-tuned version of the Qwen series optimized for Customer Relationship Management (CRM) tasks, customer support, and business communications. These quantized versions drastically reduce memory usage and accelerate inference speeds, allowing you to run a powerful 4-billion parameter CRM assistant on consumer hardware, laptops, and edge devices. This repository provides GGUF format files optimized for inference using llama.cpp, Ollama, LM Studio, and other compatible software. Note: Memory Required includes the model weight footprint plus a standard context window buffer. To run the GGUF…

Model · Image and text to text

Qwen3.5-9B-B2-Fable-Agent-OneJev-Holodeck-Lounge-q8-hi-mlx

Gheorghe Chesler

This model is a merge of: - nightmedia/Qwen3.5-9B-B2-Fable-Agent-OneJev - nightmedia/Qwen3.5-9B-Holodeck-Lounge - schneewolflabs/B2-9B - microsoft/Fara1.5-9B - armand0e/Qwen3.5-9B-Agent - armand0e/Qwen3.5-9B-Fable-5-v1 - Jackrong/Qwopus3.5-9B-Coder - DavidAU/Qwen3.5-9B-Claude-4.6-OS-Auto-Variable-HERETIC-UNCENSORED-THINKING-X8b - DavidAU/Qwen3.5-9B-GBO-Fire-HERETIC-UNCENSORED-THINKING-X8 - DavidAU/Qwen3.5-9B-Deckard-Uncensored-Heretic-Thinking - DavidAU/Qwen3.5-9B-Polaris-PolarisQwen-3NMDST2 - DavidAU/Qwen3.5-9B-Haskell-Rust-Python - DavidAU/Qwen3.5-9B-The-Bradbury-F451-Pro-Writer-Uncensored-Heretic - DavidAU/Qwen3.5-9B-Pro-Writer-1984-Orwell-Uncensored-Heretic…

Model · Image and text to text

Qwen3.5-9B-Brainwaves-Roasteramus-mxfp8-mlx

Gheorghe Chesler

This is an experimental merge between: - axiomofmind/Roasteramus - nightmedia/Qwen3.5-9B-Brainwaves I figured it might be good at code debugging. Conclusion: God's language is likely Haskell. Pure, functional, and immutable. It forces you to think about what your code means before it runs. Brainwaves (MLX) Qwen3.5-9B-Brainwaves Qwen3.5-9B (Instruct) Contribute to NightmediaAI Nightmedia is an independent AI lab located in Montana, USA. Our lab is one Macbook Pro 128GB and a few memory cards. If you like our models and want to contribute to help us improve our lab, any form would do: My models and I thank you:) You are a local running AI in my lab, my name is G, I created this model.…

This qwen35 model was trained 2x faster with Unsloth and Huggingface's TRL library.

This qwen35 model was trained 2x faster with Unsloth and Huggingface's TRL library.

This qwen35 model was trained 2x faster with Unsloth and Huggingface's TRL library.

On a laptop running Debian 13 Linux, with an Intel Core Ultra 9 185H, its integrated Arc GPU and 32 GB of memory, Janas-LLM reads a prompt at 111 tok/s and writes at 14 tok/s, 24 tok/s with its own draft block in a chat: llama.cpp, on the same laptop, 78 and 8.7. Measured on 9-11 October 2026: llama-bench -p 1024 -n 16 and janas-bench --prompt 1024 --gen 16, both at their best configuration; the drafts with janas-bench on a chat reply, the mean of three rounds (a draft guesses the next tokens and the model checks them: the same text, faster). On other machines the figures will differ. Details: Janas README. - Debian 13 (trixie) Linux, x86-64, kernel 6.12.107 - its integrated Intel Arc GPU…

Qwen3.5-9B post-trained for the Polish history matura (extended level) during the Warsaw Model Trainers alpha=64) on top of Qwen/Qwen3.5-9B. closedem = exact match on closed items. Graded scores (deepseek-flash-high@think via benchmark/bench.py) are reported separately.

Qwen3.5-9B post-trained for the Polish history matura (extended level) during the Warsaw Model Trainers alpha=64) on top of Qwen/Qwen3.5-9B. closedem = exact match on closed items. Graded scores (deepseek-flash-high@think via benchmark/bench.py) are reported separately.