Qwen3-4B-Instruct-2507

Home/Models/Qwen/Qwen3-4B-Instruct-2507

Qwen3-4B-Instruct-2507

Qwen/Qwen3-4B-Instruct-2507

kind: original
quant: bf16 (16-bit)
family: qwen
modality: text-generation
root model: Qwen/Qwen3-4B-Instruct-2507
provider: Qwen
license: apache-2.0
language: —
context: 262,144
library: transformers
vRAM (GB): —
size on disk: —
downloads: 3,506,956
likes: 905
created: —
last modified: 2025-09-17T06:56:53+00:00 (2025)
huggingface: https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507
tags: transformerssafetensorsqwen3text-generationconversationalarxiv:2505.09388license:apache-2.0eval-resultstext-generation-inferenceendpoints_compatibledeploy:azureregion:usdeploy:sagemaker

Benchmark scores (9)

benchmark	category	score	context	conf	evidence
MMLU	knowledge	80.2	MMLU-Redux	medium	`\| MMLU-Redux \| 80.2 \| 84.1 \| 77.3 \| 84.2 \|`
IFEval	instruction-following	74.5	IFEval	medium	`\| IFEval \| 74.5 \| 83.7 \| 81.2 \| 83.4 \|`
MMLU-Pro	knowledge	62.8	MMLU-Pro	medium	`\| MMLU-Pro \| 62.8 \| 69.1 \| 58.0 \| 69.6 \|`
MMLU-Pro	knowledge	56.2	MMLU-ProX	medium	`\| MMLU-ProX \| 56.2 \| 65.1 \| 49.6 \| 61.6 \|`
BFCL (function calling)	agentic	53.0	BFCL-v3	medium	`\| BFCL-v3 \| 53.0 \| 58.6 \| 57.6 \| 61.9 \|`
GPQA	reasoning	50.3	GPQA	medium	`\| GPQA \| 50.3 \| 54.8 \| 41.7 \| 62.0 \|`
LiveCodeBench	coding	31.5	LiveCodeBench v6 (25.02-25.05)	medium	`\| LiveCodeBench v6 (25.02-25.05) \| 31.5 \| 29.0 \| 26.4 \| 35.1 \|`
AIME 2025	math	22.7	AIME25	medium	`\| AIME25 \| 22.7 \| 21.6 \| 19.1 \| 47.4 \|`
Chatbot Arena	chat	15.9	Arena-Hard v2*	medium	`\| Arena-Hard v2* \| 15.9 \| 24.8 \| 9.5 \| 43.4 \|`

Discussions (29)

title	author	status	type
Qwen3 4B Instruct 2507 — the best mobile model of 2026?	3morixd	open	disc
Listed on OpenModelMap	duola15	open	disc
TemporalMesh Transformer: 29.4 PPL at 48% compute — beats Mamba, new open-source architecture	vigneshwar234	open	disc
Opened by mistake, closing this PR.	Jennny	closed	disc
Opened by mistake, closing this PR.	Jennny	closed	disc
Add MMLU-Pro evaluation result	burtenshaw	open	disc
Update README.md	Vishva007	closed	disc
Add GPQA evaluation result	burtenshaw	open	disc
注意：此模型仅支持非思考模式，不会在其输出中生成 <think></think> 块。	Jay-v2	open	disc
Safety Audit: GAE Score 40.13% (FAIL)	GAE-Auditor	draft	disc
Add 12 evaluation results to model-index	burtenshaw	closed	disc
Discrepancy in benchmark score (BFCL-v3)	mmrbulbul	open	disc
The model doesn't know about itself	sakazakiMGJ	open	disc
Not working with structured json	shubhamapollo	closed	disc
vRAM needed ?	Ashish18110	open	disc
Please support vietnamese more and more in the models.	DuongLeVan	open	disc
Adding mention of Tinker and TRL support	clem	open	disc
<tool_call> generated even with no tools or asked for	dipta007	open	disc
Release training token stats	jquessada	open	disc
Remove `<think></think>` blocks from chat template	mamousavi	draft	disc
No other 2507 models	SipOfSpike	open	disc
Sampling parameters to tau2-bench?	lewtun	open	disc
1.7b 2507?	CHNtentes	open	disc
Training completed @ 2025-08-09 11:36:54	enzii	closed	disc
Upload folder using huggingface_hub	enzii	closed	disc
Why is <think></think> required in history messages?	giangndm	open	disc
Terrible instruction following	denisalpino	open	disc
4b model with an 84.2 MMLU-Redux score?	phil111	open	disc
when 32B?	AaronFeng753	open	disc