B

Olmo 3 (AI2)

B Tier · 7.9/10

Allen Institute for AI's fully-open frontier reasoning models -- Olmo 3 family (2025-11-20) includes 7B and 32B sizes, four variants (Base, Think, Instruct, RLZero). Apache 2.0 with fully open data + checkpoints + training logs. Olmo 3-Think 32B matches Qwen3-32B-Thinking at 6x fewer training tokens

Last updated: 2026-04-17Free tier available

Score Breakdown

6.0
Ease of Use
8.0
Output Quality
9.5
Value
8.0
Features

The Good and the Bad

What we like

  • +FULLY OPEN is a different category than 'open weights.' Olmo 3 publishes not just weights but the complete training data (Dolma dataset), checkpoints at every training step, training logs, and research papers. That level of openness is irreproducible by Llama / Qwen / DeepSeek / Granite / Gemma, none of which publish the training corpus
  • +Olmo 3-Think 32B matches Qwen3-32B-Thinking reasoning quality at roughly 6x fewer training tokens -- a meaningful efficiency result that suggests AI2's data curation is genuinely competitive with the major labs at a tiny fraction of the compute budget
  • +Four-variant release (Base, Think, Instruct, RLZero) means AI2 ships matched models for pretraining research, reasoning, chat, and RLHF-free experimentation -- the RLZero variant is rare and valuable for anyone studying how instruction tuning compares to pure pretraining
  • +Non-profit research institution (AI2 = Allen Institute for AI, founded by Paul Allen) publishing at frontier scale is a durable check against the 'only for-profit labs can reach frontier' narrative. Real ecosystem importance beyond any individual model's scores

What could be better

  • Absolute quality is not frontier-competitive for its size -- Olmo 3-Think 32B is strong but DeepSeek, Qwen 3.6, GLM-5.1, and Llama 4 all outscore it at comparable sizes on mainstream benchmarks. The value is in openness + research utility, not in beating the leaderboards
  • Research audience primarily -- AI2 is optimizing for reproducibility and science, not for consumer or production ergonomics. Fine-tuning toolkits and deployment recipes are less polished than what you get with Qwen or Llama
  • Smaller community than Llama / Qwen / DeepSeek -- fewer third-party fine-tunes, fewer production deployment case studies. AI2 does serious community engagement but at a smaller scale than the for-profit labs
  • November 2025 release means it is no longer the newest open model as of April 2026 -- Granite 4, GLM-5.1, and Arcee Trinity have all shipped since. Still relevant for research-first uses, less so for bleeding-edge quality

Pricing

Self-hosted (Apache 2.0 + fully open data)

$0
  • Apache 2.0 license on weights
  • Training data FULLY OPEN -- Dolma dataset published
  • Training checkpoints at every step published
  • Training logs / ablations / paper published
  • Four variants per size: Base, Think (reasoning), Instruct (chat), RLZero (RLHF-free)

API via partner providers

Usage-based/per 1M tokens
  • Available on OpenRouter and select hosted providers
  • Pricing roughly in line with comparable 7B / 32B open-weight models
  • Primary distribution is self-hosted via Hugging Face

System Requirements

Hardware needed to self-host. Min = smallest viable setup (usually heavy quantization). Max = full-precision / production-grade.

Model variantMinMax
Olmo 3-7B (Base / Think / Instruct / RLZero)Apache 2.0 + Dolma training data fully open6 GB VRAM Q4 (RTX 3060 / 3070)24 GB VRAM FP16
Olmo 3-32B (Base / Think / Instruct / RLZero)Olmo 3-Think 32B matches Qwen3-32B-Thinking at ~6x fewer training tokens16 GB VRAM Q4 (RTX 4090)1× H100 80 GB FP16

Known Issues

  • Olmo 3's value proposition is research transparency, not peak benchmark performance. If you are choosing an open-weight model purely on MMLU / GPQA / SWE-Bench scores, Olmo will not top the list -- DeepSeek / GLM / Qwen are stronger there. Olmo earns its place when reproducibility, training-corpus transparency, or RLZero research mattersSource: AI2 Olmo 3 technical report, Interconnects analysis · 2025-11
  • Dolma training corpus is large (~3TB). Serious reproducibility work requires significant storage + compute. Most downstream users will still fine-tune from Olmo's published checkpoints rather than re-train from raw dataSource: AI2 Dolma documentation · 2025-11

Best for

AI researchers doing reproducibility work, training-data studies, instruction-tuning research, or RLHF-free (RLZero) experimentation. Also valuable for academic institutions and non-profits that want to use an open-weight model whose provenance is fully auditable. Good as a teaching / learning model where inspecting checkpoints matters.

Not for

Production deployments chasing absolute quality -- DeepSeek V3.2, GLM-5.1, Qwen 3.6, or Llama 4 score higher on mainstream benchmarks. Also not ideal for teams that need a rich community ecosystem and many third-party fine-tunes.

Our Verdict

Olmo 3 (November 2025) is the most important open-weight release for the research community in the past year, even though it does not top the mainstream benchmarks. AI2's commitment to publishing weights + data + checkpoints + logs is irreproducible by for-profit labs, and the fact that a non-profit research institution can ship 32B-scale frontier models at all is a structurally important counterweight to Big Lab concentration. For research and educational use, Olmo is the default. For production use, pick based on quality tier -- Olmo is competitive but not leading at its size.

Sources

  • Allen Institute for AI: Olmo 3 announcement (accessed 2026-04-17)
  • Interconnects: Olmo 3, America's truly open reasoning model (accessed 2026-04-17)
  • HPCwire: AI2 announces Olmo 3 family (accessed 2026-04-17)

Alternatives to Olmo 3 (AI2)

Llama 4 (Meta) logo

Llama 4 (Meta)

Meta's open-weights flagship family -- Scout (10M context), Maverick (multimodal 400B MoE), Behemoth in preview

B
7.9/10
Free tierFrom $0
Llama 4 Scout has a 10M token context wi...Llama 4 Maverick is natively multimodal ...
Updated 2026-04-13
Mistral AI logo

Mistral AI

European AI lab with open and commercial models -- Mistral Small 4 (Mar 2026, 119B MoE Apache 2.0 unified model), Medium 3 (Apr 9 2026), and Voxtral TTS (open-source speech, Mar 2026)

B
7.5/10
Free tierFrom $0
Mistral Small 4 (March 2026) unifies the...Voxtral TTS (March 2026) fills the one g...
Updated 2026-04-16
DeepSeek logo

DeepSeek

Near-frontier reasoning for pennies on the dollar -- the open-source LLM that made Silicon Valley nervous

A
8.0/10
Free tierFrom $0
Pricing is absurdly cheap compared to GP...DeepSeek-R1 reasoning model genuinely co...
Updated 2026-04-17
Gemma 4 (Google) logo

Gemma 4 (Google)

Google DeepMind's open-weights model family -- multimodal, 256K context, runs on edge devices

A
8.3/10
Free tierFrom $0
Apache 2.0 license -- truly permissive, ...Multimodal: handles text + image input (...
Updated 2026-04-08
Qwen (Alibaba) logo

Qwen (Alibaba)

Alibaba's open-weights + API family -- Qwen 3.6-Plus (Mar 30 2026, 1M context + always-on CoT + agentic tool-use), Qwen3.5 Small (2B runs on iPhone, 9B matches 120B-class models), plus Qwen3.5-Omni native multimodal. Apache 2.0 on the open sizes

A
8.8/10
Free tierFrom $0
Qwen 3.6-Plus (launched Mar 30 2026) is ...Qwen3.5 Small (0.8B / 2B / 4B / 9B) is t...
Updated 2026-04-17
GLM / Z.ai (Zhipu AI) logo

GLM / Z.ai (Zhipu AI)

Zhipu AI's open-weights family -- GLM-5.1 (launched 2026-04-07) is 744B MoE / 40B active, topped SWE-Bench Pro at 58.4 (beating GPT-5.4 and Claude Opus 4.6), MIT licensed, 200K context. Trained entirely on 100K Huawei Ascend 910B chips -- first frontier model with zero Nvidia in the training stack

A
8.0/10
Free tierFrom $0
GLM-5.1 (2026-04-07) topped SWE-Bench Pr...First frontier model trained entirely on...
Updated 2026-04-17
Kimi K2.5 (Moonshot) logo

Kimi K2.5 (Moonshot)

Moonshot's 1T-parameter MoE open-weights flagship -- best open-source agentic coder, rivals Claude Opus 4.5

A
8.1/10
Free tierFrom $0
Frontier-tier performance -- Elo 1309 on...Beats Claude Opus 4.5 on several coding ...
Updated 2026-04-13
Nemotron (Nvidia) logo

Nemotron (Nvidia)

Nvidia's open-weights family -- hybrid Mamba-Transformer MoE architecture, optimized for efficient reasoning on Nvidia hardware

B
7.8/10
Free tierFrom $0
Hybrid Mamba-Transformer architecture dr...Nemotron 3 Super activates only 3.6B par...
Updated 2026-04-17
MiniMax M2 / M2.5 logo

MiniMax M2 / M2.5

MiniMax's open-weights frontier -- first open model to match Claude Opus 4.6 on SWE-Bench at 10-20× lower cost

A
8.4/10
Free tierFrom $0
First open-weight model to hit 80.2% on ...~10B active params during inference (out...
Updated 2026-04-13
Falcon (TII) logo

Falcon (TII)

UAE's Technology Innovation Institute open-weights family -- Falcon 3 optimized for efficient sub-10B deployment on consumer hardware

B
7.1/10
Free tierFrom $0
Apache 2.0 license -- fully permissive f...Sub-10B sizes run on any consumer GPU or...
Updated 2026-04-13

gpt-oss (OpenAI)

OpenAI's FIRST open-weight models -- gpt-oss-120b (single 80GB GPU, near parity with o4-mini on reasoning) and gpt-oss-20b (runs on 16GB edge devices). Apache 2.0. Launched 2025-08-05. gpt-oss-safeguard ships in 2026 as the safety-tuned variant

A
8.1/10
Free tierFrom $0
First-ever OpenAI open-weight release --...gpt-oss-120b approaches o4-mini on reaso...
Updated 2026-04-17

IBM Granite 4.0

IBM's enterprise-focused open-weight family -- Granite 4.0 hybrid Mamba-2 + transformer architecture (70-80% memory reduction vs pure transformer), 3B to 32B sizes, Apache 2.0. First open model family to secure ISO 42001 certification. Nano 350M runs on CPU with 8-16GB RAM. 3B Vision variant landed 2026-04-01

A
8.2/10
Free tierFrom $0
Hybrid Mamba-2 + transformer architectur...Granite 4.0 Nano (350M and 1.5B) is genu...
Updated 2026-04-17

Arcee Trinity-Large-Thinking

Arcee AI's US-made open-weight frontier reasoning model -- launched 2026-04-01. 398B total params, ~13B active. Sparse MoE (256 experts, 4 active = 1.56% routing). Apache 2.0, trained from scratch. #2 on PinchBench trailing only Claude 3.5 Opus. ~96% cheaper than Opus-4.6 on agentic tasks

A
8.1/10
Free tierFrom $0
Rare US-made frontier-tier open-weight r...Trained from scratch (not a fine-tune) a...
Updated 2026-04-17

AI21 Jamba2

AI21 Labs' hybrid SSM-Transformer (Mamba-style) open-weight family -- Jamba2 launched 2026-01-08. Two sizes: 3B dense (runs on phones / laptops) and Jamba2 Mini MoE (12B active / 52B total). Apache 2.0, 256K context, mid-trained on 500B tokens

A
8.0/10
Free tierFrom $0
Hybrid SSM-Transformer (Mamba-style) arc...Jamba2 3B dense runs realistically on iP...
Updated 2026-04-17

StepFun Step 3.5 Flash

StepFun's (China) agent-focused open-weight model -- Step 3.5 Flash launched 2026-02-01. 196B sparse MoE, ~11B active. Benchmarks slightly ahead of DeepSeek V3.2 at over 3x smaller total size. Step 3 (321B / 38B active, Apache 2.0) and Step3-VL-10B multimodal also in the family

B
7.8/10
Free tierFrom $0
Step 3.5 Flash at 196B total / 11B activ...Agent-focused tuning explicitly -- tool ...
Updated 2026-04-17

Cohere Command A

Cohere's enterprise-multilingual flagship -- 111B params, 256K context, runs on 2x H100. 23 languages. CC-BY-NC 4.0 on weights (research / non-commercial), commercial requires Cohere enterprise contract. Follow-ups: Command A Reasoning + Command A Vision

B
7.5/10
Free tierFrom $0
Best-in-class multilingual open-weight m...Runs on just 2x H100 at FP16 for the ful...
Updated 2026-04-17