by Ai2

Olmo 3.1 32B Think review, pricing and limits

Ai2's strongest fully open 32B reasoning model by its own account, built for reinforcement learning research and inspectable chain-of-thought rather than leaderboard peaks.

Best for
Reasoning RL research
Not for
Long-document analysis
  • Available
  • Open source

Should you use Olmo 3.1 32B Think?

Built from an Olmo 3 run that got 21 extra days of reinforcement learning, Olmo 3.1 32B Think is for researchers and teams who must audit or retrain a reasoning model end to end. Pick it for transparency, not peak capability: newer open models report far higher GPQA Diamond scores.

Best for

  • Reasoning RL research Ai2 publishes the SFT, DPO and RL stages, their datasets and the training code, so each stage can be rerun or changed.
  • Training-data audits Open Dolma 3 and Dolci data plus OlmoTrace let you check where behavior may have come from.
  • Self-hosted math and logic work Apache 2.0 weights run on your own hardware, with AIME 2025 and MATH scores reported by Ai2.

Not for

  • Long-document analysis The window stops at 65,536 tokens.
  • Multilingual products The model card lists English only.
  • Frontier-level agents Artificial Analysis scores it 7 on its index, and newer open models report far higher GPQA Diamond.

Olmo 3.1 32B Think pricing

Pricing model
open-weights (free to download); hosted API price not confirmed
Free tier
No
TierRate
Self-hosted weightsFree download under Apache 2.0; you pay for hardware

The weights are free under a permissive license, so the cost is the hardware to run a 32B dense model. Ai2 has not published a first-party API price. Artificial Analysis lists one host, Parasail, with a $0.00 blended price that we treat as unconfirmed. For a priced open alternative, Qwen3.8 27B is sold by Alibaba Cloud at $0.50 for input and $3.00 for output, per million tokens.

Self-hosting requirements

Self-hosted
Yes

What Olmo 3.1 32B Think can do

Olmo 3.1 32B Think is a 32.2-billion-parameter dense reasoning model from Ai2, released with Apache 2.0 weights plus its training data, code and checkpoints. Ai2 reports 78.1 on AIME 2025 and 96.2 on MATH. Artificial Analysis scores it 7 on its own index, so it suits research and auditing more than frontier work.

What goes in and what comes out

Reads

  • text

Writes

  • text

Capabilities

  • Audio inputNot supported
  • Audio outputNot supported
  • Code executionNot supported
  • Video inputNot supported
  • VisionNot supported
  • Web browsingNot supported
Reasoning modes
always-on long chain-of-thought

Key features

  • Fully open model flow

    Weights, Dolma 3 pretraining data, Dolci post-training data, training code and intermediate checkpoints are all downloadable.

  • Hybrid attention

    Most layers look only at nearby text, while every fourth layer sees the entire window, which keeps long inputs affordable.

  • Extended reinforcement learning

    Ai2 kept training with reinforcement learning after the Olmo 3 launch, which it credits for a gain of more than 20 points on IFBench.

  • OlmoTrace

    In the Ai2 Playground, parts of an answer can be traced to matching documents in the training data.

  • Permissively licensed weights

    A single checkpoint that serves with Transformers, vLLM or SGLang, with no usage-threshold clauses.

Known limitations

  • The 65,536-token window is a quarter of the 262,144 tokens on Qwen3.8 27B and Gemma 4 31B.
  • The card lists English only, and the model is text in, text out.

Who it is for

  • ML researchers studying reinforcement learning for reasoning
  • Auditors and regulated teams that must document training provenance
  • Educators and students learning how LLMs are built
  • Teams that self-host and want an Apache 2.0 reasoning model

Supported languages

  • English

Olmo 3.1 32B Think benchmarks

Olmo 3.1 32B Think against its nearest rivals

GPQA Diamond % correct, higher is better

Standout result

78.1%

AIME 2025, up from 72.5% for Olmo 3 32B Think before the extra reinforcement learning

Source: Ai2 Olmo 3.1 32B Think Hugging Face model card, opened 2026-10-03

Where it sits

MeasureThis modelRankMedian
GPQA Diamond% correct57.1%#58 / 63peer median 87.8%

Ranks are against GA models on Hokai that publish the same figure; ties share a rank.

Benchmark scores

Show all 6 benchmark scores
  • MATH vendor-reported · 03 Oct 202696.2%Competition maths problems, % solved.
  • MMLU vendor-reported · 03 Oct 202686.4%General-knowledge exam across 57 subjects, % correct.
  • AIME 2025 vendor-reported · 03 Oct 202678.1%Competition-level maths problems from the 2025 exam, % solved.
  • IFBench vendor-reported · 03 Oct 202668.1%How precisely the model follows detailed instructions, % passing.
  • GPQA Diamond independent · 03 Oct 202657.1%PhD-level science questions that are hard to search for, % correct.
  • AA Intelligence Index cited: Artificial Analysis · 03 Oct 20267Composite of the evaluations Artificial Analysis runs for its index, 0 to 100.

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Safety benchmarks

Ai2 safety suite composite
83.6

Strengths and limits

Strengths

  • Best fit when the goal is research or audit: every stage is reproducible, which most open-weight releases of this size do not offer.
  • Permissive license with no usage-threshold clauses, and a dense design that is easier to serve than a mixture of experts.
  • Honest vendor documentation: the technical report discloses decontamination work and per-stage results.
  • Ai2 reports 78.1 on AIME 2025, 7.2 points above Qwen 3 32B in the same Ai2 table (vendor-reported).
Show 2 more
  • Training data, code and intermediate checkpoints are published alongside Apache 2.0 weights.
  • Reasoning traces are visible, and OlmoTrace can link parts of an answer back to training documents.

Limits

  • Low on independent measures: Artificial Analysis scores it 7 on its Intelligence Index v4.3.2.
  • Short window next to newer open models, and text-only input.
  • Hosted access is uncertain, so most teams must run their own GPUs.

Olmo 3.1 32B Think alternatives

ModelLabContextIn / out per 1MReads and writes
Olmo 3.1 32B Think This pageAi266K--text → text
Qwen3.8-27BPick Olmo 3.1 32B Think for published training data; pick Qwen3.8 27B for a much longer window and a far higher vendor-reported GPQA Diamond score.Alibaba Cloud262K$0.50 / $3text, image, video → text
Gemma 4 31BPick Olmo to audit or retrain the pipeline; pick Gemma 4 31B for image input and a longer window.Google262K--text, image → text
Phi-4Pick Olmo for a longer window and open training data; pick Phi-4 for a smaller 14B footprint under an MIT license.Microsoft16K$0.125 / $0.50text → text

Olmo 3.1 32B Think specs

Context window
66K tokens
Architecture
Transformer (dense, decoder-only, sliding-window attention with full attention every fourth layer)
Mixture of experts
No
Parameters
32.2B (32,233,522,176 parameters, dense)
VRAM requirements
The BF16 checkpoint is 64.5 GB, so the weights alone need an 80 GB accelerator or a multi-GPU split, plus KV cache.

Training data

Pretrained on Dolma 3 (about 9.3T tokens of web, olmOCR-processed science PDFs, code, math and encyclopedic text; a 5.9T-token mix), mid-trained on 100B Dolmino tokens, extended to long context on about 50B Longmino tokens, then post-trained with Dolci SFT, DPO and RLVR data. Olmo 3.1 adds 21 more days of RL on 224 GPUs.

Weight formats

  • BF16 safetensors

API endpoints

NameEndpointAuth
Weights on Hugging Facehttps://huggingface.co/allenai/Olmo-3.1-32B-ThinkNone (public download)
Self-hosted vLLM (OpenAI-compatible server)https://huggingface.co/allenai/Olmo-3.1-32B-Think?local-app=vllmNone (your own server)

SDKs

  • Python

Prompting patterns

Show 1 prompting pattern
  • Use the recommended sampling settings

    temperature 0.6, top_p 0.95, max_tokens 32768

    Ai2 evaluated the model with these settings and recommends them.

Known quirks

Show 4 known quirks
  • Degrades at low sampling temperature

    Ai2's report says both Olmo 3 Think models degenerate quickly at the low temperatures used for Olmo 2 evaluations.

    Workaround: Set temperature to 0.6 and top_p to 0.95, as the model card recommends.

  • Thinks at length before answering

    The card recommends a 32,768-token output budget, which is half of the 65,536-token window.

    Workaround: Reserve output tokens in your request and keep prompts well under half the window.

  • Vendor GPQA figure differs by source

    Ai2's technical report lists 56.7 and its model card lists 57.5, while a community record shows 57.07 on GPQA Diamond.

    Workaround: Quote the source with the number, or rerun GPQA yourself before relying on it.

  • English only

    The model card lists English as the language, and multilingual quality is not documented.

    Workaround: Test your target language first, or pick a multilingual model for production.

Olmo 3.1 32B Think facts

Licence
Apache License 2.0 · Open source
Safety posture
Balanced
Released
12 Dec 2025
Current version
Olmo-3.1-32B-Think
Knowledge cutoff
01 Dec 2024
Family
Olmo 3
Variant
Chat
Deployment
Self-hosted · Hugging Face · vLLM · SGLang · Ai2 Playground
Data retention
Self-hosted copies send no data to Ai2. We found no model-specific retention policy for the hosted Ai2 Playground.

Content restrictions

  • Ai2 says the model is intended for research and educational use under its Responsible Use Guidelines.
  • The model card warns the model can be prompted into harmful content without extra safety filtering.

Safety controls

Rlhf
No
Red teaming
no third-party partner disclosed
System prompt
Yes
Constitutional AI
No
Content filter API
No
Moderation endpoint
No

Data governance

Data residency options
self-hosted

Version history

  • Olmo-3.1-32B-Think · 12 Dec 2025

    Resumed RL for 21 more days on 224 GPUs. Ai2 reports 5+ points on AIME and 20+ on IFBench over Olmo 3 32B Think.

  • Olmo-3-32B-Think · 20 Nov 2025

    Original Olmo 3 release with Base, Think, Instruct and RL Zero models.

What changed

  • Release Olmo 3.1 32B Think released with 21 more days of reinforcement learning source
  • Release Olmo 3 launches with fully open Base, Think, Instruct and RL Zero models (Olmo-3-32B-Think) source

Questions about Olmo 3.1 32B Think

What does Olmo 3.1 32B Think cost to run?

Downloading the weights is free, so the real cost is hardware: the files total 64.5 GB. Ai2 has not published a first-party API price, and hosted access is patchy: OpenRouter listed no live endpoints on October 3, 2026, and Artificial Analysis showed a single host with a $0.00 blended price that we could not confirm. Budget for your own GPUs or a rented node rather than a per-token bill.

How does Olmo 3.1 32B Think compare with Qwen 3 32B and newer open models?

In Ai2's own table it leads Qwen 3 32B by 7.2 points on AIME and 30.8 points on IFBench, a test Ai2 wrote, but trails by about 10 points on GPQA and 6.9 on LiveCodeBench v3. Newer open models are well ahead on science questions: Qwen3.8 27B reports 89.2 on GPQA Diamond and Gemma 4 31B reports 85.7, both vendor-reported. Pick Olmo when training transparency matters more than the highest score.

Is Olmo 3.1 32B Think open source?

The weights carry a permissive Apache license, and Ai2 also releases the training data, code and intermediate checkpoints, which is what it means by fully open. The datasets are downloadable without license restrictions, according to Ai2's launch post. The model card also asks that use follow Ai2's Responsible Use Guidelines, which frame the release around research and teaching and are worth reading before a commercial launch.

Does Olmo 3.1 32B Think train on your data?

Self-hosted copies send nothing back to Ai2, so your prompts stay on your own machines. We found no model-specific retention policy for Ai2's hosted Playground, and the site privacy notice covers website data and predates this model's release. If your prompts are sensitive, run the weights locally.

Who should use Olmo 3.1 32B Think, and who should pass?

It fits researchers studying reasoning-focused reinforcement learning, auditors who need to see what a model learned from, and teams that want a permissively licensed reasoning model they can fully document. Pass if you need documents longer than its window, languages other than English, or current frontier accuracy: a newer open model such as Qwen3.8 27B or Gemma 4 31B suits those better. For chat and tool use, Ai2 points to its separate Instruct model.

About Olmo 3.1 32B Think

Olmo 3.1 32B Think is the reasoning variant of Ai2's Olmo 3 family, announced on December 12, 2025 as an update to Olmo 3 32B Think from November 2025. Ai2 resumed the reinforcement learning run for 21 more days on 224 GPUs with extra passes over its Dolci-Think-RL data, and reports gains of 5 or more points on AIME, 4 or more on ZebraLogic and IFEval, and 20 or more on IFBench. It is a dense, decoder-only Transformer with 32,233,522,176 parameters, 64 layers, a 5,120-wide hidden state and 8 key-value heads shared across 40 attention heads. Three of every four layers use a 4,096-token sliding attention window and every fourth layer attends to the full context, which Ai2 extended to 65,536 tokens.

Read the full background

The model is fully open in a stricter sense than most open-weight releases. Ai2 publishes the weights under Apache 2.0, the Dolma 3 pretraining corpus (about 9.3 trillion tokens of web pages, olmOCR-processed science PDFs, code, math and encyclopedic text), a 5.9-trillion-token pretraining mix, 100 billion tokens of mid-training data, about 50 billion tokens for long-context extension, and the Dolci sets used for supervised fine-tuning, preference tuning and reinforcement learning. Pretraining used up to 1,024 H100 GPUs, and the data cutoff is December 2024. In the Ai2 Playground, the OlmoTrace tool can trace parts of a response back to matching training documents.

Ai2's model card, checked on October 3, 2026, lists 78.1 on AIME 2025, 80.6 on AIME 2024, 96.2 on MATH, 86.4 on MMLU, 83.3 on LiveCodeBench v3, 91.5 on HumanEvalPlus, 80.1 on ZebraLogic, 93.8 on IFEval and 68.1 on IFBench. In Ai2's own table it beats Qwen 3 32B on AIME 2025 (70.9) and IFBench (37.3) but trails it on GPQA (67.3), LiveCodeBench v3 (90.2) and ZebraLogic (88.3). IFBench is an instruction-following test that Ai2 created, so read that gap with care.

Independent figures are thin. Artificial Analysis scores the model 7 on its Intelligence Index v4.3.2 (checked October 3, 2026) and publishes no speed or cost data for it. Ai2's own GPQA number differs between its sources: 56.7 in the technical report and 57.5 on the model card, while a community evaluation record on the Hugging Face page lists 57.07 on GPQA Diamond, the figure stored here. Phi-4, a 14B model, reports a similar 56.1 on GPQA Diamond, while newer releases like Qwen3.8 27B report 89.2 and Gemma 4 31B reports 85.7 on the same test, each vendor-reported.

The weights are free to download. The BF16 checkpoint is 64.5 GB across 14 safetensors files, so the weights alone need an 80 GB accelerator or a multi-GPU split before any KV cache. Transformers 4.57.0 or later, vLLM and SGLang can serve it, and Hugging Face lists 25 community quantizations. Ai2 recommends temperature 0.6, top-p 0.95 and a 32,768-token output budget. Ai2 offered the model in its Playground at launch, and its December 2025 post said inference partners would host it, but on October 3, 2026 OpenRouter showed no live endpoints and Artificial Analysis listed one host, Parasail, with no verified price.

Ai2 reports a composite safety score of 83.6 for this model, against 69.0 for Qwen 3 32B in the same table. The technical report says the development safety suite includes HarmBench, DoAnythingNow, XSTest, WildGuardTest, WildJailbreak and TrustLLM-JailbreakTrigger. The model card warns that, without safety filtering, the model can be prompted into harmful content, and the license note says it is intended for research and educational use under Ai2's Responsible Use Guidelines. No third-party red-team partner is named.

Choose it to study or reproduce reasoning-focused reinforcement learning, to audit what a model was trained on, or to self-host an Apache 2.0 reasoning model whose lineage you can document. Skip it for long documents (65,536 tokens against 262,144 on both Qwen3.8 27B and Gemma 4 31B, or 1,000,000 on NVIDIA's Nemotron 3.5 Lightning), for non-English work (the card lists English) and for agent products that need current frontier accuracy. Ai2's Olmo 3.1 32B Instruct targets chat and tool use instead. We found no newer Olmo language model as of October 3, 2026, and Ai2's Olmo co-lead Hanna Hajishirzi was reported in March 2026 to be joining Microsoft, so treat future updates as uncertain. More choices sit in the open-source models directory, and the guides Which LLM Should You Use? and Best AI Models You Can Run Locally cover picking and self-hosting.

Sources and guides

More AI Models on Hokai

  • Hy4 flashTencent's Hy4 flash, released 16 November 2026, is a lower-cost multimodal model that reads text and images across a 1M-token window.
  • Nemotron-Labs-SigmaNVIDIA's Nemotron-Labs-Sigma, released in November 2026 in 8B and 3B sizes, writes text by denoising whole blocks of tokens in parallel, not one at a time.
  • Qwen4 MaxAlibaba's Qwen4 Max, released 16 November 2026, is a sparse mixture-of-experts model with a 1M-token context and text, image and video input.
  • Gemini 4 HydrogenGoogle DeepMind's November 2026 Gemini 4 model for coding agents, with four thinking levels from low to xhigh and a free tier in AI Studio.
  • Qwen4-Max-PreviewAlibaba's November 2026 Qwen 4 flagship preview, open only to Token Plan subscribers, with three reasoning levels from low to xhigh.
  • Claude Fable 5Superseded by Fable 5.1 and Fable 5.2 but still sold on the Claude API: Anthropic's first generally available Mythos-class model, released June 2026, with a 1M-token context.