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
Score Breakdown
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)
- ✓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
- ✓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 variant | Min | Max |
|---|---|---|
| Olmo 3-7B (Base / Think / Instruct / RLZero)Apache 2.0 + Dolma training data fully open | 6 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 tokens | 16 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)
Explore more Olmo 3 (AI2) rankings
Deeper leaderboards, benchmarks, task-specific tier lists, and status/pricing pages for Olmo 3 (AI2).
The Tier List Tuesday
Weekly newsletter: tier movers, new entrants, and the VS of the week. Built from our daily AI-tool sweeps. No spam, unsubscribe anytime.
Alternatives to Olmo 3 (AI2)
Llama 4 (Meta)
Meta's open-weights family -- Scout (10M context), Maverick (multimodal 400B MoE). NOTE: Meta's frontier work moved to the proprietary Muse Spark line in April 2026; Llama remains downloadable and supported but is effectively in maintenance mode
Mistral AI
European AI lab with open and commercial models -- Le Chat is now **Vibe** (May 28 2026): one agent across Work Mode + Code Mode with a VS Code extension and CLI, powered by Mistral Medium 3.5 (128B dense, 256k context, 77.6% SWE-Bench Verified). Newest release: **Shieldstral 1.0** (Aug 4 2026), a 3B Apache 2.0 multimodal safety classifier that runs on one 16GB GPU. Earlier 2026 line: Small 4 (119B MoE Apache 2.0), Voxtral TTS. **Mistral Medium 3.1 and Medium 3 were both retired 2026-08-31** -- Medium 3.5 is the vendor-listed replacement for each
DeepSeek
DeepSeek V4 shipped 2026-04-24: V4-Pro (1.6T/49B active MoE) + V4-Flash (284B/13B active), 1M native context, Hybrid Attention Architecture, open-source on HF. **V4-Pro reached GA 2026-08-13** (Terminal Bench 2.1 87.9, three thinking-effort levels, native Responses API). **Pricing changed 2026-08-16 and the new rates are now live:** peak/off-peak billing raised rates at every hour of the day -- V4-Pro went from $0.435/$0.87 to **$0.66/$1.98 off-peak and $1.32/$3.96 peak**. Despite the 'off-peak' framing there is no hour at which it is cheaper than before
Gemma 4 (Google)
Google DeepMind's open-weights model family -- multimodal, 256K context, runs on edge devices
DiffusionGemma (Google)
Google DeepMind's experimental open-weights TEXT-DIFFUSION model (June 10, 2026) -- 26B MoE (3.8B active), Apache 2.0, generates 256-token blocks in parallel with bidirectional attention for up to 4x faster output (1,000+ tok/s on H100). Trades some quality vs Gemma 4 for raw speed
Qwen (Alibaba)
Alibaba's open-weights + API family -- Qwen3.8-Max flagship previewed at WAIC (Jul 19 2026: 2.4T sparse-MoE multimodal, closed preview, 'second only to Fable 5'), Qwen 3.7 Max GA (SWE-Bench Pro 60.6%, Terminal-Bench 69.7%, $2.50/$7.50 per 1M), Qwen3.6-27B dense Apache 2.0 (beats the 397B MoE on coding from one consumer GPU)
GLM / Z.ai (Zhipu AI)
Zhipu AI's open-weights flagship -- GLM-5.2 (launched 2026-06-13) is a ~753B-parameter MoE with a 1M-token context and the new IndexShare sparse-attention architecture (~2.9x lower per-token FLOPs at 1M context), MIT licensed. Vendor benchmarks put SWE-Bench Pro at 62.1 (up from GLM-5.1's 58.4) and it tops the Artificial Analysis open-weights Intelligence Index; VentureBeat reports it beats GPT-5.5 on several long-horizon coding benchmarks at roughly 1/6 the cost. Drop-in for Claude Code / Cline / OpenCode. Still trained outside the Nvidia stack on Huawei Ascend silicon
Kimi K3 (Moonshot)
Moonshot's 2.8T-parameter Kimi K3 (launched 2026-07-16/17) is the largest open-weight model ever released -- 1M context, multimodal, $3/$15 per 1M via API, ranked best-available on Arena.AI at launch. WEIGHTS SHIPPED ~2026-07-26/27 on Hugging Face (2.8T total / 104B activated, safetensors) under a custom Kimi K3 License, not the Modified MIT of the K2 line
Nemotron (Nvidia)
Nvidia's open-weights family -- hybrid Mamba-Transformer MoE architecture, optimized for efficient reasoning on Nvidia hardware. Nemotron 3 Ultra (550B total / 55B active) shipped 2026-06-04 as the family flagship, joining Super (120B/12B, March) and Nano
MiniMax M3
MiniMax's coding/agent flagship -- M3 (June 1 2026): 1M-token context, MSA sparse attention (>15x decoding speedup at long context), SWE-Bench Pro 59.0%, Terminal-Bench 66.0%. OPEN WEIGHTS LIVE on HuggingFace since June 12 (~428B total / ~23B active, native multimodal, minimax-community license)
Falcon (TII)
UAE's Technology Innovation Institute open-weights family -- Falcon 3 optimized for efficient sub-10B deployment on consumer hardware
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
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
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
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
StepFun Step 3.7 Flash
StepFun's (China) agent-focused open-weight family -- Step 3.7 Flash (May 28 2026): 198B sparse MoE vision-language model, ~11B active, 256K context, Apache 2.0, ~400 tok/s, SWE-Bench Pro 56.3. Supersedes Step 3.5 Flash (Feb 2026) as the flagship
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
LongCat-2.0 (Meituan)
Meituan's open-source 1.6T-parameter MoE (~48B active) with native 1M-token context, MIT license -- trained entirely on domestic Chinese AI ASICs and revealed as the stealth 'Owl Alpha' model that had been topping OpenRouter
Inkling (Thinking Machines Lab)
Mira Murati's $12B lab ships its first model (2026-07-15): a 975B/41B-active open-weights MoE that reasons natively over text, images, and audio with a 1M-token context -- positioned not as the strongest model, but as the best starting point for fine-tuning via Tinker
Bonsai 27B (PrismML)
The first 27B-class model that runs on a phone (2026-07-14) -- ternary and 1-bit quantizations of Qwen3.6 27B squeeze a multimodal, tool-calling, 262K-context model into 3.9-5.9GB under Apache 2.0