Qwen3-4B-Thinking-2507

4.02B parameters, text in, text out, 256K tokens of context. 2 providers serve it from $0.010 per million input tokens.

Verified 2026-10-25 Rank #102 in LLMs · score 99 method v1 Sources 3 Downloads 30d 661.6k · #142 of 619 tracked

Changed this week: 1 new quantized build on the Hub (144 in all).2026-10-24 · 2 new providers: Featherless, Nscale.2026-10-23

Can I run it?2.4GB at INT4Estimatefits a 16 GB cardmethodCommunity: Run Qwen3-4B-Thinking-2507 Locally | Atomic Chat →

What does it cost?$0.010/1M incheapest of 1Nscale2026-10-25

Can I use it?Apache 2.0Permissive; patent grantlicence file2026-10-24

Do people use it?661.6kdownloads, 30 days#142 of 619 we track · top 23%+3 likes, downloads +9% this week627 likesHugging Face2026-10-24

Worth reading

Picked by rule from 18 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Most discussed · 490 HN pts

GPT-OSS vs. Qwen3 and a detailed look how things evolved since GPT-2

OpenAI released gpt-oss-120b and gpt-oss-20b, their first open-weight models since GPT-2 in 2019.

A 2025 research paper confirms Dropout results in worse downstream performance in single-epoch regimes.

magazine.sebastianraschka.com 2025-08-10 490 HN pts · 97 comments 5.2k words about the Qwen3 family family

Latest · 8 months ago

Qwen3.5 122B and 35B models offer Sonnet 4.5 performance on local computers

venturebeat.com 2026-02-28 461 HN pts · 272 comments about the Qwen3 family family

18 best of 18 tracked

The 5 picks above, plus these 13. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

ArticleQwen3-Coder-Nextqwen.ai · 2026-02-03 · official
ArticleQwen3-Max-Thinkingqwen.ai · 2026-01-26 · official
ArticleQwen3-TTS family is now open sourced: Voice design, clone, and generationqwen.ai · 2026-01-22 · official
ArticleQwen3: Think deeper, act fasterqwenlm.github.io · 2025-04-28 · 869 HN pts · official
ArticleQwen3-Coder: Agentic coding in the worldqwenlm.github.io · 2025-07-22 · official
ArticleQwen3-Nextqwen.ai · 2025-09-12 · official
ArticleQwen3-VLqwen.ai · 2025-09-23 · official
CodeQwen3-Omni: Native Omni AI model for text, image and videogithub.com · 2025-09-22
ArticleQwen3-4B-Thinking: “This is art—pelicans don't ride bikes!”simonwillison.net · undated
ArticleQwen3-4B-Thinking-2507: Efficient Chain-of-Thoughtemergentmind.com · undated
Article🚀 Qwen3-4B-Thinking-2507 just shipped!dev.to · undated
ArticleQwen3-4B-Thinking-2507: Pricing & Benchmarks | LLMLearnerllmlearner.com · undated
ArticleAlibaba Qwen Unveils Qwen3-4B-Instruct-2507 and Qwen3-4B-Thinking-2507: Refreshing the Importance of Small Language Models – The TechBriefsthetechbriefs.com · undated

Show all 13

Where to run it

Live hosted endpoints, cheapest first, USD per million tokens. Ties break on output price, then provider name.

#	Cheapest 2	In /1M	Out /1M	Context	Quant	Seen
1	Nscale	$0.010	$0.030	256K	—	Oct 25
2	Featherless	—	—	—	—	Oct 25

All 2 providers →

Where it's used

Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.

Served by 2 API providers; heavily run locally (144 quantized builds); a big base for builders (367 fine-tunes, adapters and merges).

As a service2Featherless, Nscale
On own hardware144quantized builds
Built on367272 fine-tunes · 71 adapters · 24 merges
In apps63Hugging Face Spaces

Top Spaces running it: tomg-group-umd/lm-watermarking ♥ 100 · dqy08/InfoLens1 ♥ 22 · prithivMLmods/NAVA-Text-to-Video ♥ 11 · pesu-dev/askpesu ♥ 6

Will it fit?

Weights only, an estimate: leave room for the KV cache. Method

Fits a 16 GB card at INT4

About 2.4 GB at INT4, 4.8 GB at FP8, 9.7 GB at BF16: weights only, so leave room for the KV cache.

Estimate params × bytes per param × 1.2 · how we estimate

The story so far

Weights, every dated article, and repo commits on one rail.

Most coverage came in the first two weeks

Teal is the weights (2025-08-05), amber triangles are 3 dated articles (each links to it), grey ticks are repo commits.

What people built on it

511 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.

511 derivatives on the Hub, most of them fine-tunes

From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.

GGUF MaziyarPanahi/Qwen3-4B-Thinking-2507-GGUF for llama.cpp, Ollama, LM Studio 651.6k downloads MLX lmstudio-community/Qwen3-4B-Thinking-2507-MLX-4bit for Apple silicon 2.5M downloads FP8 Qwen/Qwen3-4B-Thinking-2507-FP8 for vLLM, SGLang 2.4M downloads

All variants →

Reference

Everything else, folded: the full tables, every source.

Benchmarks

24 rows

Benchmarks 24 rows

measured third-party leaderboard rowauthor-reported number from the model card or eval files (24)No confidence intervals unless the source gave them; settings differ between rows, so compare within a benchmark, not across. Date = the benchmark's own date when it gives one, else when we read it.

Benchmark	Value	Setting / notes	Reported by	Date
author-reported AIME 2025 · Reason through hard problems	81.3	Qwen3-4B-Thinking-2507	model author	2026-10-07
author-reported Arena-Hard · Chat and follow instructions	34.9	Qwen3-4B-Thinking-2507	model author	2026-10-07
author-reported BFCL · Call tools and APIs	71.2	Qwen3-4B-Thinking-2507	model author	2026-10-07
author-reported CFEval	1852	Qwen3-4B-Thinking-2507	model author	2026-10-07
author-reported Creative Writing · Chat and follow instructions	75.6	Qwen3-4B-Thinking-2507	model author	2026-10-07
author-reported Creative Writing · Chat and follow instructions	83.3	Qwen3-4B-Thinking-2507	model author	2026-10-07
author-reported GPQA Diamond · Reason through hard problems	65.8	Qwen3-4B-Thinking-2507	model author	2026-10-07
author-reported HMMT · Reason through hard problems	55.5	Qwen3-4B-Thinking-2507	model author	2026-10-07
author-reported IFEval · Chat and follow instructions	87.4	Qwen3-4B-Thinking-2507	model author	2026-10-07
author-reported INCLUDE · Work across languages	64.4	Qwen3-4B-Thinking-2507	model author	2026-10-07
author-reported LiveBench · Reason through hard problems	71.8	Qwen3-4B-Thinking-2507	model author	2026-10-07
author-reported LiveCodeBench · Write code with agents	55.2	Qwen3-4B-Thinking-2507	model author	2026-10-07
author-reported MMLU · Reason through hard problems	86.1	Qwen3-4B-Thinking-2507	model author	2026-10-07
author-reported MMLU-Pro · Reason through hard problems	74	Qwen3-4B-Thinking-2507	model author	2026-10-07
author-reported MMLU-ProX · Work across languages	64.2	Qwen3-4B-Thinking-2507	model author	2026-10-07
author-reported Multi-IF · Chat and follow instructions	77.3	Qwen3-4B-Thinking-2507	model author	2026-10-07
author-reported OJBench	17.9	Qwen3-4B-Thinking-2507	model author	2026-10-07
author-reported PolyMATH · Work across languages	46.2	Qwen3-4B-Thinking-2507	model author	2026-10-07
author-reported SuperGPQA · Reason through hard problems	47.8	Qwen3-4B-Thinking-2507	model author	2026-10-07
author-reported TAU1-Airline	48	Qwen3-4B-Thinking-2507	model author	2026-10-07
author-reported TAU1-Retail	66.1	Qwen3-4B-Thinking-2507	model author	2026-10-07
author-reported tau-bench · Call tools and APIs	53.5	Qwen3-4B-Thinking-2507	model author	2026-10-07
author-reported tau-bench · Call tools and APIs	58	Qwen3-4B-Thinking-2507	model author	2026-10-07
author-reported tau-bench · Call tools and APIs	27.2	Qwen3-4B-Thinking-2507	model author	2026-10-07

What the card says

1 cited sentence

Summary generated, cited

The model features a 262,144 native context length and is released under the Apache 2.0 license.^[1]^[2]

“262,144 natively”
license

Generated 2026-10-07 15:02 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.

Use-case evidence

7 use cases

What it is good at 7 use cases

Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.

Facts

21 fields

Listed in LLMs

Facts

Qwen3-4B-Thinking-2507 has 4.02B parameters. (safetensors metadata, observed 2026-10-24)
Context length is 262,144 tokens (256K). (config.json max_position_embeddings, observed 2026-10-24)
License: Apache 2.0 (friction: none). Permissive; patent grant. (license file, observed 2026-10-24)
Created 2025-08-05, last modified 2025-08-06. (HF API, observed 2026-10-24)

Pipeline tag	text-generation
Library	transformers
Architectures	Qwen3ForCausalLM
Model type	qwen3
Dtype	bfloat16
Params by dtype	BF16: 4,022,468,096
Size band	M
MoE	no
Context length	262,144 (config.json max_position_embeddings)
Papers	arXiv:2505.09388
License	Apache 2.0 (apache-2.0)
Hub snapshot	2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · Qwen3ForCausalLM · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · Qwen3ForCausalLM · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed

vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.

Hardware fit

3 estimates

Hardware fit ESTIMATE

Weights-only memory for Qwen3-4B-Thinking-2507: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)	9.7 GB · Single GPU (<24 GB)
FP8 (1 byte/param)	4.8 GB · Single GPU (<24 GB)
INT4 (0.5 bytes/param)	2.4 GB · Edge (<4 GB)

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra.

All variants

40 variants

Lineage

Original model Family: Qwen/Qwen3 511 derivatives on the Hub

finetune: 272quantized: 144merge: 24adapter: 71

Top variants (40 shown, ranked by downloads)

Variant	Relation	Quant	Params	Band	License	Author	Likes	Downloads
prefeitura-rio/Rio-3.0-Open-Mini	finetune	—	—	—	mit	prefeitura-rio	9	3.9M
lmstudio-community/Qwen3-4B-Thinking-2507-MLX-4bit	quantized	MLX	4.02B	M	apache-2.0	lmstudio-community	14	2.5M
lmstudio-community/Qwen3-4B-Thinking-2507-MLX-8bit	quantized	MLX	4.02B	M	apache-2.0	lmstudio-community	8	2.5M
lmstudio-community/Qwen3-4B-Thinking-2507-MLX-6bit	quantized	MLX	4.02B	M	apache-2.0	lmstudio-community	2	2.5M
Qwen/Qwen3-4B-Thinking-2507-FP8	quantized	-FP8	4.41B	M	apache-2.0	Qwen	75	2.4M
MaziyarPanahi/Qwen3-4B-Thinking-2507-GGUF	quantized	GGUF	—	—	—	MaziyarPanahi	2	651.6k
cyankiwi/Qwen3-4B-Thinking-2507-AWQ-8bit	quantized	AWQ	1.80B	S	apache-2.0	cyankiwi	4	492.3k
GeneralAnalysis/GA_Guard_Thinking	finetune	—	4.02B	M	cc-by-nc-4.0	GeneralAnalysis	4	475.5k
lmstudio-community/Qwen3-4B-Thinking-2507-GGUF	quantized	GGUF	—	—	—	lmstudio-community	26	335.6k
unsloth/Qwen3-4B-Thinking-2507-GGUF	quantized	GGUF	—	—	apache-2.0	unsloth	108	237.1k
unsloth/Qwen3-4B-Thinking-2507-unsloth-bnb-4bit	quantized	-bnb	4.12B	M	apache-2.0	unsloth	2	164.7k
janhq/Jan-v1-4B	finetune	—	4.02B	M	apache-2.0	janhq	353	140.3k
PantheonUnbound/Satyr-V0.1-4B	quantized	GGUF	—	—	apache-2.0	PantheonUnbound	242	114.4k
unsloth/Qwen3-4B-Thinking-2507	finetune	—	4.02B	M	apache-2.0	unsloth	8	109.5k
AllThingsIntel/Apollo-V0.1-4B-Thinking	quantized	GGUF	4.02B	M	other	AllThingsIntel	41	88.4k
bartowski/Qwen_Qwen3-4B-Thinking-2507-GGUF	quantized	GGUF	—	—	—	bartowski	19	68.2k
mlx-community/Qwen3-4B-Thinking-2507-4bit	quantized	-4bit	4.02B	M	apache-2.0	mlx-community	4	29.7k
openbmb/AgentCPM-Explore	finetune	—	4.02B	M	apache-2.0	openbmb	418	8k
adeelahmad/ReasonableQwen3-4B	quantized	GGUF	4.02B	M	apache-2.0	adeelahmad	6	20.3k
Alcoft/Qwen_Qwen3-4B-Thinking-2507-GGUF	quantized	GGUF	—	—	apache-2.0	Alcoft	0	18k
JunHowie/Qwen3-4B-Thinking-2507-GPTQ-Int4	quantized	GPTQ	4.02B	M	apache-2.0	JunHowie	1	16k
sequelbox/Qwen3-4B-Thinking-2507-DAG-Reasoning	finetune	—	4.02B	M	apache-2.0	sequelbox	8	10.2k
unsloth/Qwen3-4B-Thinking-2507-bnb-4bit	quantized	-bnb	4.14B	M	apache-2.0	unsloth	2	9k
Ma7ee7/Qwen3.8_4B_Distilled	finetune	—	4.02B	M	apache-2.0	Ma7ee7	8	8.7k
khazarai/Qwen3-4B-Gemini-3.1-Pro-Reasoning-Distilled	quantized	GGUF	—	—	apache-2.0	khazarai	2	8.9k
cyankiwi/Qwen3-4B-Thinking-2507-AWQ-4bit	quantized	AWQ	4.41B	M	apache-2.0	cyankiwi	5	8.7k
detrax/Qwen3-4B-Thinking-2507-Qwen3.8-Max-Distillation-Detrax	quantized	GGUF	—	—	apache-2.0	detrax	4	8.5k
huihui-ai/Huihui-Qwen3-4B-Thinking-2507-abliterated	finetune	—	4.02B	M	apache-2.0	huihui-ai	26	6.9k
Mungert/DMind-3-mini-GGUF	quantized	GGUF	—	—	apache-2.0	Mungert	1	8k
giladgd/Qwen3-4B-Thinking-2507-GGUF	quantized	GGUF	—	—	apache-2.0	giladgd	0	7.2k
qvac/MedPsy-4B	finetune	—	4.41B	M	apache-2.0	qvac	15	5.8k
DripGhad1223/Qwen3-4B-Kimi2.5-Reasoning-Distilled-GGUF	quantized	GGUF	—	—	apache-2.0	DripGhad1223	4	6k
mradermacher/Qwen3-4B-Thinking-2507-i1-GGUF	quantized	GGUF	—	—	apache-2.0	mradermacher	0	6k
abcorrea/bw-v1	finetune	—	4.02B	M	—	abcorrea	0	6k
abcorrea/sok-v3	finetune	—	4.02B	M	—	abcorrea	0	5.7k
abcorrea/sok-v5	finetune	—	4.02B	M	—	abcorrea	0	5.7k
prithivMLmods/Qwen3-4B-Thinking-2507-GGUF	quantized	GGUF	—	—	apache-2.0	prithivMLmods	1	5.2k
ggml-org/Qwen3-4B-Thinking-2507-Q8_0-GGUF	quantized	Q8_0	—	—	apache-2.0	ggml-org	2	5k
Mungert/Jan-v1-4B-GGUF	quantized	GGUF	—	—	apache-2.0	Mungert	0	5.1k
DevQuasar/Qwen.Qwen3-4B-Thinking-2507-GGUF	quantized	GGUF	—	—	—	DevQuasar	0	5k

Repository changelog

7 commits

changed in the last 7 daysNew providers: 2026-10-23Other: 2026-10-24all changes

Repository changelog 7 commits

Date	Commit	SHA
2025-08-06	Update README.md	768f209
2025-08-06	Delete configuration.json	cd97167
2025-08-06	Delete .mv	2e95fd2
2025-08-06	Delete .msc	51853d6
2025-08-06	Upload folder using huggingface_hub	fe0b03e
2025-08-05	Upload folder using huggingface_hub	a847ed9
2025-08-05	initial commit	4000519

All providers

2 providers · 2 endpoints

Where to run 2 providers · 2 endpoints

Hosted endpoints observed for Qwen3-4B-Thinking-2507, sorted by input price (USD per million tokens). Each row names its observer and date; prices change daily, so check the source before quoting. Cheapest observed input price: $0.01/M.

Provider	Model ref	Context	Quant	$/M in	$/M out	tok/s	TTFT	Tools	Status	Observed
Nscale hf	Qwen/Qwen3-4B-Thinking-2507	256K	—	$0.01	$0.03	160.9	270 ms	yes	live	2026-10-25
Featherless hf	Qwen/Qwen3-4B-Thinking-2507	—	—	—	—	—	—	—	live	2026-10-25

Articles and coverage

18 pieces

Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.

Official1

Qwen3: Think deeper, act faster

qwenlm.github.io2025-04-28officialauthorityHN front pageabout the Qwen3 familyHN 869 pts · 388 comments2,139 words

Qwen3 introduces Hybrid Thinking Modes, allowing users to toggle between step-by-step reasoning and rapid responses via computational budget control. The open-weight MoE model Qwen3-235B-A22B achieves performance similar to dense models while using only 10% of the active parameters. Pre-training utilized approximately 36 trillion tokens across 119 languages, with deployment supported by SGLang and vLLM.written, reviewed, cited

Qwen3-8B and Qwen3-14B models do not use tied embeddings, unlike smaller variants.
The pre-training pipeline includes a final stage extending context length to 32K tokens.

Analysis9

Qwen3.5 122B and 35B models offer Sonnet 4.5 performance on local computers

venturebeat.com2026-02-28HN front pageabout the Qwen3 familyHN 461 pts · 272 comments

GPT-OSS vs. Qwen3 and a detailed look how things evolved since GPT-2

magazine.sebastianraschka.com2025-08-10authorityHN front pageabout the Qwen3 familyHN 490 pts · 97 comments5,195 words

OpenAI released gpt-oss-120b and gpt-oss-20b, their first open-weight models since GPT-2 in 2019. The 20B model can run on a consumer GPU with up to 16 GB of RAM, and the 120B model can run on a single H100 with 80 GB of RAM or newer hardware. The models utilize MXFP4 optimization to fit onto single GPUs.written, reviewed, cited

A 2025 research paper confirms Dropout results in worse downstream performance in single-epoch regimes.

Qwen3-4B-Thinking: “This is art—pelicans don't ride bikes!”

simonwillison.netauthority

Qwen3 4B 2507 - Intelligence, Performance & Price Analysis

artificialanalysis.aiauthority

Qwen3-4B-Thinking-2507: Efficient Chain-of-Thought

emergentmind.com

Qwen3 4B Thinking 2507: benchmarks, context window — ModelCap #136

modelcap.ai

Run Qwen3-4B-Thinking-2507 Locally | Atomic Chat

atomic.chat

🚀 Qwen3-4B-Thinking-2507 just shipped!

dev.to

Qwen3-4B-Thinking-2507: Pricing & Benchmarks | LLMLearner

llmlearner.com

All coverage tracked for this model (18) — append-only, nothing is ever removed

Title	Source	Published	First seen	Signals
Qwen3-4B-Thinking: “This is art—pelicans don't ride bikes!”	simonwillison.net	—	2026-10-07	exa
Qwen3 4B 2507 - Intelligence, Performance & Price Analysis	artificialanalysis.ai	—	2026-10-07	exa
Qwen3-4B-Thinking-2507: Efficient Chain-of-Thought	emergentmind.com	—	2026-10-07	exa
Qwen3 4B Thinking 2507: benchmarks, context window — ModelCap #136	modelcap.ai	—	2026-10-07	exa
Run Qwen3-4B-Thinking-2507 Locally \| Atomic Chat	atomic.chat	—	2026-10-07	exa
🚀 Qwen3-4B-Thinking-2507 just shipped!	dev.to	—	2026-10-07	exa
Qwen3-4B-Thinking-2507: Pricing & Benchmarks \| LLMLearner	llmlearner.com	—	2026-10-07	exa
Alibaba Qwen Unveils Qwen3-4B-Instruct-2507 and Qwen3-4B-Thinking-2507: Refreshing the Importance of Small Language Models – The TechBriefs	thetechbriefs.com	—	2026-10-07	exa
Qwen3.5 122B and 35B models offer Sonnet 4.5 performance on local computers	venturebeat.com	2026-02-28	2026-10-07	hn
Qwen3-Coder-Next	qwen.ai official	2026-02-03	2026-10-07	hn
Qwen3-Max-Thinking	qwen.ai official	2026-01-26	2026-10-07	hn
Qwen3-TTS family is now open sourced: Voice design, clone, and generation	qwen.ai official	2026-01-22	2026-10-07	hn
Qwen3-VL	qwen.ai official	2025-09-23	2026-10-07	hn
Qwen3-Omni: Native Omni AI model for text, image and video	github.com	2025-09-22	2026-10-07	hn
Qwen3-Next	qwen.ai official	2025-09-12	2026-10-07	hn
GPT-OSS vs. Qwen3 and a detailed look how things evolved since GPT-2	magazine.sebastianraschka.com	2025-08-10	2026-10-07	hn
Qwen3-Coder: Agentic coding in the world	qwenlm.github.io official	2025-07-22	2026-10-07	hn
Qwen3: Think deeper, act faster	qwenlm.github.io official	2025-04-28	2026-10-07	hn

Papers

1 paper

Papers 1

arXiv links found in the model card and its tags. "Peer-reviewed" is shown only with evidence: a published venue in OpenAlex (measured) or the authors' own journal reference on arXiv (reported).

Qwen3 Technical Report this model

An Yang, Anfeng Li, Baosong Yang +57 · 2025-05-14 · PDF · HF paper page (349 upvotes)

Similar models

3 models

Similar models

Phi-3-mini-4k-instruct · 3.82B
same size band (M), same task (text-generation), fewer providers, permissive license
Phi-3.5-mini-instruct · 3.82B
same size band (M), same task (text-generation), permissive license
Phi-4-mini-instruct · 3.84B
same size band (M), same task (text-generation), permissive license

Sources

3 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:35 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:45 UTC · Summary: 2026-10-07 15:02 UTC · Last verified 2026-10-25

Follow Qwen3-4B-Thinking-2507

Watch this model

/feeds/models/Qwen--Qwen3-4B-Thinking-2507.xml(RSS 2.0) · every change to Qwen3-4B-Thinking-2507
/feeds/families/Qwen--Qwen3.xml(RSS 2.0) · every change across the Qwen3 family

Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.