Olmo-3-7B-Think

7.30B parameters, text in, text out, 64K tokens of context.

Verified 2026-10-25 Rank #674 in LLMs · score 72.6 method v1 Sources 3 Downloads 30d 487.6k · #166 of 619 tracked

Changed this week: 1 new quantized build on the Hub (34 in all).2026-10-24

Can I run it?4.4GB at INT4Estimatefits a 16 GB cardmethod

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

Do people use it?487.6kdownloads, 30 days#166 of 619 we track · top 27%+0 likes, downloads +0% this week110 likesHugging Face2026-10-24

Worth reading

Picked by rule from 9 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Latest · 26 days ago

Reproducing Olmo 3 7B Pre-training

Google Cloud reproduced the Allen Institute for AI's Olmo 3 7B model using MaxText on TPUs, matching the original training loss curve and held-out metrics.

The team lost three quarters of capacity mid-run but resumed on a quarter-sized slice with no recipe change.

developers.googleblog.com 2026-09-29 1 HN pts 5.2k words

9 best of 9 tracked

The 1 pick above, plus these 8. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

PaperNCP-ArchPreview: 8.9B latent LM matches OLMo-3-7B on 51% of tokensarxiv.org · 2026-09-14 · 2 HN pts
ArticleOlmo 3: Charting a path through the model flow to lead open-source AIallenai.org · 2025-11-20 · 390 HN pts · official
ArticleOlmo 3 is a fully open LLMsimonwillison.net · 2025-11-23 · 5 HN pts
ArticleOlmo 3: America's open reasoning modelsinterconnects.ai · 2025-11-20 · 4 HN pts
CodeOlmo 3 from Scratch (Standalone Notebook)github.com · 2025-11-24 · 1 HN pts
ArticleOlmo 3 and the Open LLM Renaissancecameronrwolfe.substack.com · 2025-12-15 · 4 HN pts
DiscussionAi2 Olmo 3, a new SOTA open LLM (7B and 32B)twitter.com · 2025-11-20 · 2 HN pts
DiscussionWe present Olmo 3, our next family of open, leading language modelstwitter.com · 2025-11-20 · 1 HN pts

Show all 8

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.

No API provider serves it, so its downloads are people running it themselves: 34 quantized builds and 8 Spaces.

As a service0no API provider listed
On own hardware34quantized builds
Built on15817 fine-tunes · 140 adapters · 1 merges
In apps8Hugging Face Spaces

Will it fit?

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

Fits a 16 GB card at INT4

About 4.4 GB at INT4, 8.8 GB at FP8, 18 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.

Coverage peaked at launch and is still arriving

Teal is the weights (2025-11-18), amber triangles are 9 dated articles (each links to it), grey ticks are repo commits.

What people built on it

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

192 derivatives on the Hub, most of them adapters

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

GGUF unsloth/Olmo-3-7B-Think-GGUF for llama.cpp, Ollama, LM Studio 54.1k downloads MLX lmstudio-community/Olmo-3-7B-Think-MLX-8bit for Apple silicon 1.7k downloads AWQ kaitchup/Olmo-3-7B-Think-awq-w4a16-asym for vLLM, SGLang 232 downloads

All variants →

Reference

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

Benchmarks

31 rows

Benchmarks 31 rows

measured third-party leaderboard rowauthor-reported number from the model card or eval files (31)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 AGI Eval	77.2	Olmo 3 Think 7B SFT	model author	2026-10-07
author-reported AGI Eval	79.1	Olmo 3 Think 7B DPO	model author	2026-10-07
author-reported AGI Eval	81.5	Olmo 3 Think 7B	model author	2026-10-07
author-reported AIME 2024 · Reason through hard problems	69.6	Olmo 3 Think 7B SFT	model author	2026-10-07
author-reported AIME 2024 · Reason through hard problems	74.6	Olmo 3 Think 7B DPO	model author	2026-10-07
author-reported AIME 2024 · Reason through hard problems	71.6	Olmo 3 Think 7B	model author	2026-10-07
author-reported AIME 2025 · Reason through hard problems	57.6	Olmo 3 Think 7B SFT	model author	2026-10-07
author-reported AIME 2025 · Reason through hard problems	62.7	Olmo 3 Think 7B DPO	model author	2026-10-07
author-reported AIME 2025 · Reason through hard problems	64.6	Olmo 3 Think 7B	model author	2026-10-07
author-reported BBH · Reason through hard problems	84.1	Olmo 3 Think 7B SFT	model author	2026-10-07
author-reported BBH · Reason through hard problems	83.7	Olmo 3 Think 7B DPO	model author	2026-10-07
author-reported BBH · Reason through hard problems	86.6	Olmo 3 Think 7B	model author	2026-10-07
author-reported HumanEval · Write code with agents	88.2	Olmo 3 Think 7B SFT	model author	2026-10-07
author-reported HumanEval · Write code with agents	91.4	Olmo 3 Think 7B DPO	model author	2026-10-07
author-reported HumanEval · Write code with agents	89.9	Olmo 3 Think 7B	model author	2026-10-07
author-reported IFEval · Chat and follow instructions	77.9	Olmo 3 Think 7B SFT	model author	2026-10-07
author-reported LCB v3	67.8	Olmo 3 Think 7B SFT	model author	2026-10-07
author-reported LCB v3	75.1	Olmo 3 Think 7B DPO	model author	2026-10-07
author-reported LCB v3	75.2	Olmo 3 Think 7B	model author	2026-10-07
author-reported MATH-500 · Reason through hard problems	94.4	Olmo 3 Think 7B SFT	model author	2026-10-07
author-reported MATH-500 · Reason through hard problems	92.4	Olmo 3 Think 7B DPO	model author	2026-10-07
author-reported MATH-500 · Reason through hard problems	95.1	Olmo 3 Think 7B	model author	2026-10-07
author-reported MBPP · Write code with agents	63.2	Olmo 3 Think 7B SFT	model author	2026-10-07
author-reported MBPP · Write code with agents	63	Olmo 3 Think 7B DPO	model author	2026-10-07
author-reported MBPP · Write code with agents	64.7	Olmo 3 Think 7B	model author	2026-10-07

Show all 31 rows (6 more)

Benchmark	Value	Setting / notes	Reported by	Date
author-reported OMEGA	45	Olmo 3 Think 7B SFT	model author	2026-10-07
author-reported OMEGA	40.5	Olmo 3 Think 7B DPO	model author	2026-10-07
author-reported OMEGA	37.8	Olmo 3 Think 7B	model author	2026-10-07
author-reported ZebraLogic	57.9	Olmo 3 Think 7B SFT	model author	2026-10-07
author-reported ZebraLogic	60.6	Olmo 3 Think 7B DPO	model author	2026-10-07
author-reported ZebraLogic	66.5	Olmo 3 Think 7B	model author	2026-10-07

What the card says

1 cited sentence

Summary generated, cited

This finetune uses long chain-of-thought thinking to improve reasoning tasks like math and coding.^[1]^[2]

lineage kind
“Long chain-of-thought thinking improves reasoning tasks l…”

Generated 2026-10-07 15:14 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.

Use-case evidence

5 use cases

What it is good at 5 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.

Reason through hard problemsauthor-reportedAuthor reports MATH 95.1 Author reports AIME 2024 74.6 Author reports AIME 2025 64.6 +3 more
Chat and follow instructionsauthor-reportedAuthor reports IFEval 77.9 tag: conversational chat template present
Write code with agentsauthor-reportedAuthor reports HumanEval+ 91.4 Author reports MBPP+ 64.7
Run on a laptop or phonestructuraldecision-model estimate p=0.97 fits: single-gpu
Call tools and APIsstructuralchat template: functions.

Facts

21 fields

Listed in LLMs

Facts

Olmo-3-7B-Think has 7.30B parameters. (safetensors metadata, observed 2026-10-24)
Context length is 65,536 tokens (64K). (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 2025-11-18, last modified 2026-06-25. (HF API, observed 2026-10-24)

Pipeline tag	text-generation
Library	transformers
Architectures	Olmo3ForCausalLM
Model type	olmo3
Dtype	bfloat16
Params by dtype	BF16: 7,298,011,136
Size band	M
MoE	no
Context length	65,536 (config.json max_position_embeddings)
Languages	en
Papers	arXiv:2512.13961
License	Apache 2.0 (apache-2.0)
Hub snapshot	2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · Olmo3ForCausalLM · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: not in registry · 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 Olmo-3-7B-Think: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)	18 GB · Single GPU (<24 GB)
FP8 (1 byte/param)	8.8 GB · Single GPU (<24 GB)
INT4 (0.5 bytes/param)	4.4 GB · Single GPU (<24 GB)

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

All variants

40 variants

Lineage

Fine-tuned from allenai/Olmo-3-7B-Think-DPO (same organization) Family: allenai/Olmo-3 192 derivatives on the Hub

Base model: allenai/Olmo-3-7B-Think-DPO

quantized: 34finetune: 17adapter: 140merge: 1

Top variants (40 shown, ranked by downloads)

Variant	Relation	Quant	Params	Band	License	Author	Likes	Downloads
unsloth/Olmo-3-7B-Think-GGUF	quantized	GGUF	—	—	apache-2.0	unsloth	10	54.1k
mradermacher/Olmo-3-7B-Think-i1-GGUF	quantized	GGUF	—	—	apache-2.0	mradermacher	1	10.1k
bartowski/allenai_Olmo-3-7B-Think-GGUF	quantized	GGUF	—	—	apache-2.0	bartowski	0	8.9k
eaddario/Olmo-3-7B-Think-GGUF	quantized	GGUF	—	—	apache-2.0	eaddario	0	5.2k
unsloth/Olmo-3-7B-Think	finetune	—	528K	S	apache-2.0	unsloth	2	4.1k
lmstudio-community/Olmo-3-7B-Think-GGUF	quantized	GGUF	—	—	apache-2.0	lmstudio-community	2	4k
unsloth/Olmo-3-7B-Think-unsloth-bnb-4bit	quantized	-bnb	7.50B	M	apache-2.0	unsloth	2	3k
mradermacher/Olmo-3-7B-Think-GGUF	quantized	GGUF	—	—	apache-2.0	mradermacher	0	2.7k
DevQuasar/allenai.Olmo-3-7B-Think-GGUF	quantized	GGUF	—	—	—	DevQuasar	0	1.9k
lmstudio-community/Olmo-3-7B-Think-MLX-8bit	quantized	MLX	7.30B	M	apache-2.0	lmstudio-community	1	1.7k
MaziyarPanahi/Olmo-3-7B-Think-GGUF	quantized	GGUF	—	—	—	MaziyarPanahi	0	1.1k
Tuwhy/Olmo-3-7B-Think-OPSA	finetune	—	7.30B	M	apache-2.0	Tuwhy	1	932
richardyoung/Olmo-3-7B-Think-heretic	finetune	—	7.30B	M	apache-2.0	richardyoung	1	873
lmstudio-community/Olmo-3-7B-Think-MLX-4bit	quantized	MLX	7.30B	M	apache-2.0	lmstudio-community	0	724
lmstudio-community/Olmo-3-7B-Think-MLX-6bit	quantized	MLX	7.30B	M	apache-2.0	lmstudio-community	0	553
mlx-community/Olmo-3-7B-Think-4bit	quantized	-4bit	7.30B	M	apache-2.0	mlx-community	0	459
firmanda/Olmo-3-7B-Think-GGUF	quantized	GGUF	—	—	apache-2.0	firmanda	0	451
AS-SiliconMind/SiliconMind-V1-Olmo-3-7B-Think	finetune	—	7.30B	M	apache-2.0	AS-SiliconMind	1	390
AIML-TUDA/Olmo-3.1-7B-Think	finetune	—	528K	S	apache-2.0	AIML-TUDA	2	336
mlx-community/Olmo-3-7B-Think-8bit	quantized	-8bit	7.30B	M	apache-2.0	mlx-community	1	296
FBTMAML/Olmo-3-1125-7B-Think	quantized	GGUF	—	—	apache-2.0	FBTMAML	0	340
jkim96/Olmo-3-7B-Think-DASHQ-Q2-GGUF	quantized	GGUF	—	—	apache-2.0	jkim96	0	329
johnsmith968530/allenai-Olmo-3-7B-Think-MLX-8bit	quantized	MLX	7.30B	M	apache-2.0	johnsmith968530	0	308
mattritchey/Olmo-3-7B-Think-Q4_K_M-GGUF	quantized	GGUF	—	—	apache-2.0	mattritchey	0	269
dams2005/olmo-3-7b-think-saes	finetune	—	—	—	—	dams2005	2	145
0labs-in/Sky-7B	finetune	—	21.2B	L	apache-2.0	0labs-in	0	240
kaitchup/Olmo-3-7B-Think-awq-w4a16-asym	quantized	awq	1.69B	S	apache-2.0	kaitchup	0	232
Samll/olmo3-7b-think-thinkprune-b4000	finetune	—	7.30B	M	—	Samll	0	230
NikolayKozloff/Olmo-3-7B-Think-Q8_0-GGUF	quantized	Q8_0	—	—	apache-2.0	NikolayKozloff	1	157
CCATresearch/Olmo-3-7B-Thinking	quantized	GGUF	—	—	apache-2.0	CCATresearch	0	183
mlx-community/Olmo-3-7B-Think-bfloat16	finetune	—	7.30B	M	apache-2.0	mlx-community	0	177
introvoyz041/Olmo-3-7B-Think-mlx-4Bit	quantized	mlx	7.30B	M	apache-2.0	introvoyz041	0	174
AXONVERTEX-AI-RESEARCH/Olmo-3-7B-Think-Q8_0-GGUF	quantized	Q8_0	—	—	apache-2.0	AXONVERTEX-AI-RESEARCH	0	166
0labs-in/Sky-21B	finetune	—	21.2B	L	apache-2.0	0labs-in	0	166
kaitchup/Olmo-3-7B-Think-fp8-dynamic	quantized	-fp8	7.30B	M	apache-2.0	kaitchup	0	147
jondale/Olmo-3-7B-Think-GGUF	quantized	GGUF	—	—	apache-2.0	jondale	0	143
mlx-community/Olmo-3-7B-Think-6bit	quantized	—	7.30B	M	apache-2.0	mlx-community	0	118
kaitchup/Olmo-3-7B-Think-w8a8-smoothquant	quantized	-w8a8	7.30B	M	apache-2.0	kaitchup	0	112
kaitchup/Olmo-3-7B-Think-gptq-w4a16-g128	quantized	gptq	7.35B	M	apache-2.0	kaitchup	0	109
litert-community/Olmo-3-7B-Think	quantized	—	—	—	apache-2.0	litert-community	1	53

Repository changelog

25 commits

changed in the last 7 daysOther: 2026-10-24all changes

Repository changelog 25 commits

Date	Commit	SHA
2026-06-25	Update the number of parameters (#7)	d97e442
2026-01-05	Add OLMo 3 citation	7c991fd
2026-01-05	Add OLMo 3 citation	c4d93a2
2025-12-10	Update config.json (#5)	61748f8
2025-12-10	Define "do_sample" explicitly in generation_config.json (#6)	5d89684
2025-11-24	Update README.md	b3744c9
2025-11-24	Update README.md	520872c
2025-11-24	Upload olmo-think.png	aa94ba7
2025-11-24	Update README.md	2d9c0ad
2025-11-21	Update README.md	53d899c
2025-11-21	Update README.md	b9cbb32
2025-11-21	Update README.md	27ffc20
2025-11-21	Update generation_config.json	032a160
2025-11-21	Add `transformers` as the library name (#1)	a39e3f8
2025-11-20	Update README.md	fd68d46
2025-11-20	Update README.md	5608d10
2025-11-20	Update README.md	e0f48d4
2025-11-20	Update README.md	1bb3407
2025-11-20	Update README.md	f6a52c1
2025-11-20	Update README.md	74b0721
2025-11-20	Update README.md	d4fc39b
2025-11-20	Create README.md	a155986
2025-10-31	Update config.json	923f0a3
2025-10-31	Upload folder using huggingface_hub	361b459
2025-10-31	initial commit	64dab40

Articles and coverage

9 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.

Official2

Reproducing Olmo 3 7B Pre-training

developers.googleblog.com2026-09-29authorityHN 1 pts · 0 comments5,194 words

Google Cloud reproduced the Allen Institute for AI's Olmo 3 7B model using MaxText on TPUs, matching the original training loss curve and held-out metrics. The team converted the PyTorch architecture to JAX, implementing a reordered-norm block, QK-norm, and 3:1 sliding/global attention while catching a data-loader bug that caused memorization. The run tracked the full ~5.93T-token / 1.41M-step budget, sustaining 57.4% MFU on v5p devices during stage 2.written, reviewed, cited

The team lost three quarters of capacity mid-run but resumed on a quarter-sized slice with no recipe change.

Olmo 3: Charting a path through the model flow to lead open-source AI

allenai.org2025-11-20officialauthorityHN front pageabout the Olmo 3 familyHN 390 pts · 125 comments3,773 words

AllenAI released Olmo 3.1 Think 32B and Olmo 3.1 Instruct 32B, the most performant checkpoints to date. The Think variant extended the best reinforcement learning run with an additional 21 days on 224 GPUs, yielding improvements of 5+ points on AIME, 4+ points on ZebraLogic, 4+ points on IFEval, and 20+ points on IFBench. The Instruct variant applies the recipe behind Olmo 3 Instruct 7B to a larger model for chat and tool use.written, reviewed, cited

Olmo 3-Think (32B) narrows the gap to top open models while training on roughly 6x fewer tokens than competitors.
The model flow allows users to swap in domain-specific data during mid-training or adjust post-training for specific use cases.
Olmo 3-Base (32B) achieves competitive performance with Qwen 2.5 and Gemma 3 while maintaining full openness of weights and code.

Analysis5

NCP-ArchPreview: 8.9B latent LM matches OLMo-3-7B on 51% of tokens

arxiv.org2026-09-14authorityHN 2 pts · 0 comments436 words

NCP-ArchPreview is a latent-space language model that pushes autoregressive pretraining beyond standard next-token prediction by learning to predict discrete concepts spanning multiple tokens. The 8.9B parameter model trains on 5.73T tokens from the Dolma-3 dataset, consuming only 51.3% of total training tokens to match OLMo-3-7B's loss while outperforming it by 2.45 points on the downstream macro-average and 5.99 points on GSM8K. Updating just the 17M-parameter VQ module yields a novel, lightweight interface for domain adaptation.written, reviewed, cited

Olmo 3 is a fully open LLM

simonwillison.net2025-11-23authorityabout the Olmo 3 familyHN 5 pts · 2 comments1,220 words

Olmo 3 is a fully open LLM series from Ai2 that releases training data, processes, and checkpoints alongside model weights. The 32B-scale model is pretrained on a ~9.3-trillion-token corpus and trains on roughly 6x fewer tokens than competing models. During testing, the 32B Think model took 14 minutes 43 seconds to generate 8,437 tokens for a simple SVG request.written, reviewed, cited

The 7B Instruct model download size is 4.16GB.
OlmoTrace currently relies on phrase matching to find training documents.

Olmo 3: America's open reasoning models

interconnects.ai2025-11-20authorityabout the Olmo 3 familyHN 4 pts · 0 comments1,814 words

Olmo 3 introduces a new family of fully open language models available today, featuring 7B and 32B variants. This family represents the first 32B (or larger) fully open reasoning model and the best 7B Western-origin thinking and instruct models. The 32B Thinker scores get within 1-2 points overall with their respective Qwen3 8/32B models.written, reviewed, cited

Olmo 3 Instruct models represent the best 7B scale model from a Western or American company.
The 32B range is largely devoid of non-Qwen competition for reasoning models.

Olmo 3 from Scratch (Standalone Notebook)

github.com2025-11-24authorityabout the Olmo 3 familyHN 1 pts · 0 comments323 words

Olmo 3 and the Open LLM Renaissance

cameronrwolfe.substack.com2025-12-15about the Olmo 3 familyHN 4 pts · 0 comments13,013 words

Community2

Ai2 Olmo 3, a new SOTA open LLM (7B and 32B)

twitter.com2025-11-20about the Olmo 3 familyHN 2 pts · 0 comments

We present Olmo 3, our next family of open, leading language models

twitter.com2025-11-20about the Olmo 3 familyHN 1 pts · 0 comments

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).

Olmo 3 this model

Team Olmo, Allyson Ettinger, Amanda Bertsch +65 · 2025-12-15 · PDF · HF paper page (38 upvotes)

Similar models

3 models

Similar models

Llama-3.1-8B-Instruct · 8.03B
same size band (M), same task (text-generation), more providers
Qwen2.5-7B-Instruct · 7.62B
same size band (M), same task (text-generation), more providers, permissive license
Mistral-7B-Instruct-v0.2 · 7.24B
same size band (M), same task (text-generation), more providers, permissive license

Sources

3 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:42 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:14 UTC · Last verified 2026-10-25

Follow Olmo-3-7B-Think

Watch this model

/feeds/models/allenai--Olmo-3-7B-Think.xml(RSS 2.0) · every change to Olmo-3-7B-Think
/feeds/families/allenai--Olmo-3.xml(RSS 2.0) · every change across the Olmo-3 family

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