# Qwen3.5-4B by Qwen: Open-Weight Model
Source: https://savrn.com/models/qwen3-5-4b
Markdown alternate of the page above; the site index is https://savrn.com/llms.txt

---

## Runs On

What it takes to serve Qwen3.5-4B (4.7B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.

| Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits |
| --- | --- | --- | --- | --- | --- |
| 16-bit | 9.3 GB | 11.2 GB | 1x [MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) (192 GB) Vultr | $1.85 | [1x H100](https://savrn.com/ai-index/pricing/gpus/h100) $1.99 · [1x MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) $2.00 |
| 8-bit | 4.7 GB | 5.6 GB | 1x [MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) (192 GB) Vultr | $1.85 | [1x H100](https://savrn.com/ai-index/pricing/gpus/h100) $1.99 · [1x MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) $2.00 |
| 4-bit | 2.3 GB | 2.8 GB | 1x [MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) (192 GB) Vultr | $1.85 | [1x H100](https://savrn.com/ai-index/pricing/gpus/h100) $1.99 · [1x MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) $2.00 |

Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the [SAVRN Index](https://savrn.com/ai-index/pricing/gpus), read Oct 25, 2026.

[Qwen3.5-4B on every accelerator the SAVRN Index prices, at every precision](https://savrn.com/models/qwen3-5-4b/gpus)

## SAVRN's Notes on Qwen3.5-4B

4.7B parameters that read images and text over a 262,144-token window: 11.2 GB at 16-bit on the one 192 GB MI300X at $1.85 an hour that anchors our small-model pricing, 5.6 GB at 8-bit, 2.8 GB at 4-bit for several copies per card. Quantizing here is a packing choice, not a requirement. First published 2026-02-27; the hub records 7.4M downloads a month.

Apache 2.0 clears commercial use, fine-tuning and redistribution. It derives from Qwen3.5-4B-Base, and the page lists an official Qwen3.5-4B-GGUF plus community MLX 4-bit conversions for Apple silicon. Reported evaluations — GPQA Diamond 76.2, MMMU-Pro vision 66.3, MMLU-Pro 79.1 — are strong for the size, and the long derived list, from NuExtract3 to motion planners, shows how heavily the community fine-tunes this substrate.

SAVRN Research

## Model Card

By Qwen, published under apache-2.0, revision 851bf6e806ef.

[!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:

[Read the full model card (3,179 words)](https://savrn.com/models/qwen3-5-4b/card)

## Configuration

Architecture

Qwen3_5ForConditionalGeneration

Context length (tokens)

262,144

Layers

32

Hidden size

2,560

Feed-forward size

9,216

Attention heads

16

Key/value heads

4

Head dimension

256

Vocabulary size

248,320

Model type

qwen3_5

## Identity and Version

Repository

Qwen/Qwen3.5-4B

Publisher

Qwen

Task

Image and text to text

Modality

Image and text

Library

transformers

Parameters

4.7B parameters

Languages

Not stated by the source

Revision

851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a

First published

2026-02-27

Last updated

2026-03-02

## Files and Weights

14 files, 9.3 GB in total. The weights are 2 files totalling 9.3 GB in safetensors.

Weights2 files · 9.3 GB

Configuration4 files · 80.1 KB

Tokenizer4 files · 22.9 MB

Documentation2 files · 89.2 KB

Other1 file · 7.8 KB

Repository1 file · 1.6 KB

Every file

| File | Type | Size | SHA-256 |
| --- | --- | --- | --- |
| model.safetensors-00001-of-00002.safetensors | Weights | 5.3 GB | 26a93f066e19 |
| model.safetensors-00002-of-00002.safetensors | Weights | 4.0 GB | cb544bd9bfae |
| config.json | Configuration | 3.2 KB | — |
| model.safetensors.index.json | Configuration | 76.2 KB | — |
| preprocessor_config.json | Configuration | 390 B | — |
| video_preprocessor_config.json | Configuration | 385 B | — |
| LICENSE | Documentation | 11.5 KB | — |
| README.md | Documentation | 77.7 KB | — |
| chat_template.jinja | Other | 7.8 KB | — |
| .gitattributes | Repository | 1.6 KB | — |
| merges.txt | Tokenizer | 3.4 MB | — |
| tokenizer.json | Tokenizer | 12.8 MB | 5f9e4d4901a9 |
| tokenizer_config.json | Tokenizer | 16.7 KB | — |
| vocab.json | Tokenizer | 6.7 MB | — |

## License and Download

License

apache-2.0

Access

Open weights, no gate

Download size

9.3 GB

[Download from Qwen](https://modelscope.cn/models/Qwen/Qwen3.5-4B)

Released by Qwen through ModelScope. [Read the license](https://www.apache.org/licenses/LICENSE-2.0).

## Built From

- Derived from Qwen/Qwen3.5-4B-Base

## Evaluations

Each result is shown as reported, with the conditions its reporter stated. None is a SAVRN measurement. A comparison lines two results up only when their configuration, unit and setup are all stated and identical.

| Benchmark | Conditions | Result | Reported by | Revision | Date |
| --- | --- | --- | --- | --- | --- |
| Idavidrein/gpqa | Task diamondMetric diamondComparison conditions not established | 76.2 | [Model Card](https://huggingface.co/Qwen/Qwen3.5-4B) Reported by a third party | Evaluated revision not stated | 2026-03-02 |
| LiquidAI/ifstruct-v1.0 | Task ifstruct_v1Metric ifstruct_v1Comparison conditions not established | 36.25 | Liquid AI — IFStruct v1.0 blog (Qwen3.5-4B) Reported by a third party | Evaluated revision not stated | 2026-06-30 |
| MMMU/MMMU_Pro | Task mmmu_pro_visionMetric mmmu_pro_visionComparison conditions not established | 66.3 | [Model Card](https://huggingface.co/Qwen/Qwen3.5-4B) Reported by a third party | Evaluated revision not stated | 2026-04-28 |
| TIGER-Lab/MMLU-Pro | Task mmlu_proMetric mmlu_proComparison conditions not established | 79.1 | [Model Card](https://huggingface.co/Qwen/Qwen3.5-4B) Reported by a third party | Evaluated revision not stated | 2026-03-02 |
| likaixin/ScreenSpot-Pro | Task overallMetric overallComparison conditions not established | 60.3 | [Model Card](https://huggingface.co/Qwen/Qwen3.5-4B) Reported by a third party | Evaluated revision not stated | 2026-03-18 |
| llamaindex/ExtractBench | Task longMetric longSetup Pipeline name: qwen3_5_4b_vllm_extract_oneshot_structured_output_fileComparison conditions not established | 27.31 | [ExtractBench](https://huggingface.co/datasets/llamaindex/ExtractBench) Reported by a third party | Evaluated revision not stated | 2026-08-26 |
| llamaindex/ExtractBench | Task meanMetric meanSetup Pipeline name: qwen3_5_4b_vllm_extract_oneshot_structured_output_fileComparison conditions not established | 82.43 | [ExtractBench](https://huggingface.co/datasets/llamaindex/ExtractBench) Reported by a third party | Evaluated revision not stated | 2026-08-26 |
| llamaindex/ExtractBench | Task mediumMetric mediumSetup Pipeline name: qwen3_5_4b_vllm_extract_oneshot_structured_output_fileComparison conditions not established | 76.64 | [ExtractBench](https://huggingface.co/datasets/llamaindex/ExtractBench) Reported by a third party | Evaluated revision not stated | 2026-08-26 |
| llamaindex/ExtractBench | Task shortMetric shortSetup Pipeline name: qwen3_5_4b_vllm_extract_oneshot_structured_output_fileComparison conditions not established | 89.06 | [ExtractBench](https://huggingface.co/datasets/llamaindex/ExtractBench) Reported by a third party | Evaluated revision not stated | 2026-08-26 |
| llamaindex/ParseBench | Task chartMetric chartSetup Pipeline name: qwen3_5_4b_vllm_layoutComparison conditions not established | 2.5 | [ParseBench](https://huggingface.co/datasets/llamaindex/ParseBench) Reported by a third party | Evaluated revision not stated | 2026-04-14 |
| llamaindex/ParseBench | Task layoutMetric layoutSetup Pipeline name: qwen3_5_4b_vllm_layoutComparison conditions not established | 19.7 | [ParseBench](https://huggingface.co/datasets/llamaindex/ParseBench) Reported by a third party | Evaluated revision not stated | 2026-04-14 |
| llamaindex/ParseBench | Task meanMetric meanSetup Pipeline name: qwen3_5_4b_vllm_layoutComparison conditions not established | 35.4 | [ParseBench](https://huggingface.co/datasets/llamaindex/ParseBench) Reported by a third party | Evaluated revision not stated | 2026-04-14 |
| llamaindex/ParseBench | Task tableMetric tableSetup Pipeline name: qwen3_5_4b_vllm_layoutComparison conditions not established | 8 | [ParseBench](https://huggingface.co/datasets/llamaindex/ParseBench) Reported by a third party | Evaluated revision not stated | 2026-04-14 |
| llamaindex/ParseBench | Task text_contentMetric text_contentSetup Pipeline name: qwen3_5_4b_vllm_layoutComparison conditions not established | 88.9 | [ParseBench](https://huggingface.co/datasets/llamaindex/ParseBench) Reported by a third party | Evaluated revision not stated | 2026-04-14 |
| llamaindex/ParseBench | Task text_formattingMetric text_formattingSetup Pipeline name: qwen3_5_4b_vllm_layoutComparison conditions not established | 57.8 | [ParseBench](https://huggingface.co/datasets/llamaindex/ParseBench) Reported by a third party | Evaluated revision not stated | 2026-04-14 |

## Memory Requirements

| Precision | Weights in memory |
| --- | --- |
| As published | 9.3 GB |
| 16-bit | 9.3 GB |
| 8-bit | 4.7 GB |
| 4-bit | 2.3 GB |

Weights only, from the published parameter count; the key-value cache and runtime add to this.

## Built on This Model

- Quantized from[Qwen3.5-4B-GGUF](https://savrn.com/models/qwen3-5-4b-gguf)
- Derived from[Qwen3.5-4B-GGUF](https://savrn.com/models/qwen3-5-4b-gguf)
- Derived from[NuExtract3](https://savrn.com/models/nuextract3)
- Quantized from[Qwen3.5-4B](https://savrn.com/models/qwen3-5-4b-2)
- Derived from[Qwen3.5-4B](https://savrn.com/models/qwen3-5-4b-2)
- Derived from[h2o-lightning-4b](https://savrn.com/models/h2o-lightning-4b)
- Adapter of[hopper](https://savrn.com/models/hopper)
- Derived from[hopper](https://savrn.com/models/hopper)
- Derived from[dQwen3.5-4B-Base](https://savrn.com/models/dqwen3-5-4b-base)
- Quantized from[Hunter-1-4B](https://savrn.com/models/hunter-1-4b)
- Derived from[Hunter-1-4B](https://savrn.com/models/hunter-1-4b)
- Derived from[Valen-4B](https://savrn.com/models/valen-4b)
- Derived from[blink-4b](https://savrn.com/models/blink-4b)
- Quantized from[Qwen3.5-4B-RKLLM](https://savrn.com/models/qwen3-5-4b-rkllm)
- Derived from[Qwen3.5-4B-RKLLM](https://savrn.com/models/qwen3-5-4b-rkllm)
- Derived from[messier-one](https://savrn.com/models/messier-one)
- Derived from[Decision-1.0-Nox-4B](https://savrn.com/models/decision-1-0-nox-4b)
- Derived from[ChindaMT-4B](https://savrn.com/models/chindamt-4b)
- Derived from[Vinci-Piccolo-1.0](https://savrn.com/models/vinci-piccolo-1-0)
- Derived from[P110-ThaiOCR-4B](https://savrn.com/models/p110-thaiocr-4b)
- Quantized from[Qwen3.5-4B-W4A16-AutoRound-NVFP4](https://savrn.com/models/qwen3-5-4b-w4a16-autoround-nvfp4)
- Derived from[Qwen3.5-4B-W4A16-AutoRound-NVFP4](https://savrn.com/models/qwen3-5-4b-w4a16-autoround-nvfp4)
- Derived from[aplomb-1](https://savrn.com/models/aplomb-1)
- Quantized from[Qwen3.5-4B-W4A16-GPTQ](https://savrn.com/models/qwen3-5-4b-w4a16-gptq)
- Derived from[Qwen3.5-4B-W4A16-GPTQ](https://savrn.com/models/qwen3-5-4b-w4a16-gptq)
- Derived from[Qwen3.5-4B-onwx](https://savrn.com/models/qwen3-5-4b-onwx)
- Derived from[Xiaomi-Lens-0](https://savrn.com/models/xiaomi-lens-0)
- Derived from[WildAWM-4B](https://savrn.com/models/wildawm-4b)
- Adapter of[Gleipnir-4B](https://savrn.com/models/gleipnir-4b)
- Derived from[Gleipnir-4B](https://savrn.com/models/gleipnir-4b)
- Quantized from[Qwen3.5-4B-Orinfer-E8P-Q4A8](https://savrn.com/models/qwen3-5-4b-orinfer-e8p-q4a8)
- Derived from[Qwen3.5-4B-Orinfer-E8P-Q4A8](https://savrn.com/models/qwen3-5-4b-orinfer-e8p-q4a8)
- Derived from[OpenJudgement-4B-Preview](https://savrn.com/models/openjudgement-4b-preview)
- Adapter of[checkpoints](https://savrn.com/models/checkpoints)
- Derived from[checkpoints](https://savrn.com/models/checkpoints)
- Derived from[metask-jev-4b-policy-mix](https://savrn.com/models/metask-jev-4b-policy-mix)
- Derived from[surfPRM-qwen3.5-4b](https://savrn.com/models/surfprm-qwen3-5-4b)
- Derived from[SDRPN-Qwen3.5-4B](https://savrn.com/models/sdrpn-qwen3-5-4b)
- Derived from[jevify-qwen3.5-4b-readout-coh](https://savrn.com/models/jevify-qwen3-5-4b-readout-coh)
- Derived from[jevify-qwen3.5-4b-readout](https://savrn.com/models/jevify-qwen3-5-4b-readout)

## Compare Qwen3.5-4B

- [Qwen3.5-4B vs Qwen3-VL-4B-Instruct](https://savrn.com/models/compare/qwen3-5-4b-vs-qwen3-vl-4b-instruct)
- [Qwen3.5-4B vs vllm-translategemma-4b-it](https://savrn.com/models/compare/qwen3-5-4b-vs-vllm-translategemma-4b-it)
- [gemma-3-4b-it vs Qwen3.5-4B](https://savrn.com/models/compare/gemma-3-4b-it-vs-qwen3-5-4b)

## Questions About Qwen3.5-4B

### How much GPU memory does Qwen3.5-4B need?

About 11.2 GB at 16-bit and 2.8 GB at 4-bit: the weights (4.7B parameters) plus a working margin. A long context needs more.

### What is the cheapest GPU to run Qwen3.5-4B on?

At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.

### Can I use Qwen3.5-4B commercially?

Yes. Qwen3.5-4B is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.

### What is Qwen3.5-4B's context length?

262,144 tokens, from the maximum position embeddings in its published configuration.

## Similar Models

Model · Image and text to text

### [P110-ThaiOCR-4B](https://savrn.com/models/p110-thaiocr-4b)

[Porsuk Rungarunsrisuk](https://savrn.com/model-publishers/porsukr)

P110-ThaiOCR-4B is a 4B vision-language model for Thai documents: printed and handwritten text, tables, forms, government PDFs and scans. It reads pages into plain text or structured output and can answer questions about a document (key information, classification, charts), without per-template rules. BF16 weights. Recommended prompt for plain transcription (Thai): Tested with vLLM 0.29.0 and transformers 5.17.0. All open models were run on the same hardware with the same scoring code (October 2026). Qwen3.5-4B is the base model P110 was fine-tuned from. Models with two runs are averaged. Bold marks the best score in each row. GPT-6 Luna is a closed model, called through its API. Like every…

Open weights apache-2.0 4.7B parameters 262,144 tokens transformers

[View model](https://savrn.com/models/p110-thaiocr-4b)

Model · Image and text to text

### [Lodestar-4B](https://savrn.com/models/lodestar-4b)

[StartLux](https://savrn.com/model-publishers/startlux-models)

Lodestar-4B is a 4-billion-parameter decision model. You give it a state (plain text, JSON or a long document) and one or more typed questions; it returns a probability for every listed option. Each answer is read from a single forward pass. The model never generates free text, so there is nothing to parse and no output length to budget for. (Probabilities rounded to three decimals.) The same call from Python, run inside the downloaded folder: A score question takes its levels as a list, lowest first, and returns {"type": "score", "score":, "probabilities": {"0": p0, "1": p1,...}}. Every question is rendered into one chat prompt (thinking disabled): The probability of each option is the…

Open weights apache-2.0 4.7B parameters 262,144 tokens transformers

[View model](https://savrn.com/models/lodestar-4b)

Model · Image and text to text

### [Xiaomi-Lens-0](https://savrn.com/models/xiaomi-lens-0)

[SeerRay-Lab](https://savrn.com/model-publishers/seerray-lab)

Xiaomi-Lens-0 is a 4B multimodal language model built on Qwen3.5-4B. It adds a lightweight pixel-mask decoder and is trained with joint language and mask supervision. Each generated token identifies an image region that the decoder turns into a segmentation mask. The model supports referring expression comprehension and segmentation, multi-instance grounding, reasoning segmentation, embodied referring, and fine-grained visual perception. It can produce masks in its final answer or during reasoning. This repository contains the model weights, configuration, tokenizer, processor, and chat template. Inference uses the XiaomiLensQwen35ForConditionalGeneration class from the accompanying…

Open weights apache-2.0 4.6B parameters 262,144 tokens transformers

[View model](https://savrn.com/models/xiaomi-lens-0)

Model · Image and text to text

### [Valen-4B](https://savrn.com/models/valen-4b)

[Valen](https://savrn.com/model-publishers/valen-team)

Valen-4B combines Qwen3.5-4B with a two-layer MLP-Mixer decision head. It predicts probability distributions for Choice, Noul (yes/no), and Score questions. A shared state can contain text, images, or video and serve multiple questions. Valen-4B 基于 Qwen3.5-4B 和两层 MLP-Mixer 决策头，支持图文、文本、视频，以及共享 state 的多题推理。 These weights include merged LoRA adapters. BF16 rounding can cause small probability differences from the original training checkpoints used for the reported benchmarks. The verified environment uses Python 3.10, PyTorch 2.6.0, Transformers 5.4.0, torchvision 0.21.0, Pillow, and av 16.1.0. Flash Attention 2 requires a compatible flash-attn installation. For a local copy, replace the…

Open weights apache-2.0 4.5B parameters 262,144 tokens transformers

[View model](https://savrn.com/models/valen-4b)

Model · Image and text to text

### [LT-OPD](https://savrn.com/models/lt-opd)

[Ruixuan Yang](https://savrn.com/model-publishers/yyy051007)

LT-OPD trains vision-language models with fewer visual tokens through on-policy self-distillation. This Qwen3.5-4B model combines CDPruner with an MLP that aggregates discarded tokens into their nearest retained tokens, preserving their original positions without adding tokens. Use Python 3.12 and the CUDA environment in the installation guide. Load the model with LT-OPD's visual-token compression runtime

Open weights apache-2.0 4.5B parameters 262,144 tokens transformers

[View model](https://savrn.com/models/lt-opd)

Model · Image and text to text

### [h2o-lightning-4b](https://savrn.com/models/h2o-lightning-4b)

[H2O.ai](https://savrn.com/model-publishers/h2oai)

H2O-Lightning-4B is a 4B-parameter decision model from H2O.ai, built on Qwen/Qwen3.5-4B. It answers typed decision questions about a record (a document, a ticket, a policy, a conversation) and about images that come with the record (photos, screenshots, scanned documents, charts). It returns a probability for every option: - yes/no (noul): gives the probability that a statement is true; It runs on unmodified vLLM 0.30.0 with a small standard-library shim in front (h2olightningshim.py). Each decision is one forward pass and one output token, so the cost is input tokens only. Try it in your browser: live demo on Hugging Face Spaces (text and images; built by the Hugging Face team and running…

Open weights apache-2.0 4.5B parameters 262,144 tokens transformers

[View model](https://savrn.com/models/h2o-lightning-4b)

## Qwen

[All models and datasets](https://savrn.com/model-publishers/qwen)

## Versions

- [851bf6e806ef](https://savrn.com/models/qwen3-5-4b/versions/851bf6e806ef) · current 2026-10-25

## Explore More

- [All image and text to text models](https://savrn.com/models/tasks/image-and-text-to-text)
- [All models under apache-2.0](https://savrn.com/models/licenses/apache-2-0)
- [Model comparisons](https://savrn.com/models/comparisons)
- [The model directory](https://savrn.com/models)
- [Open model prices by host](https://savrn.com/ai-index/pricing/open-models)

## Source

- Repository metadata, read 2026-10-19.
- [Qwen website](https://qwen.ai/)
- [Hugging Face record](https://huggingface.co/Qwen/Qwen3.5-4B)
- [How the hub is built](https://savrn.com/model-hub/methodology)