OLMo by AllenAI
Every step open: weights, training data, code and checkpoints
Overview
OLMo is Ai2's fully open language model family (OLMo 3, up to 32B) — open weights, training data (Dolma), training code and evals, not just a paper. Runs via Hugging Face, Ollama or vLLM; Ai2 calls the 32B model the best open-source LLM released to date.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 11 Jul 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
Use cases
What you can produce with OLMo by AllenAI
- Open weights (Apache 2.0), up to 32B parameters
- Full open training data (Dolma 2, 7T tokens)
- Open training code (OLMo-core)
- Public evaluation harness (OLMES)
- Published intermediate training checkpoints
- Post-training recipe (Tulu 3)
- Runs via Hugging Face, Ollama, vLLM
ASEAN Perspective
OLMo by AllenAI in Southeast Asia
ASEAN-region availability and pricing notes coming soon. Drop the editorial team a note via /contact/ if you can supply local context (Singapore/Malaysia/Indonesia/Thailand/Vietnam).
OLMo earns its 'radically open' label: not just weights but the full Dolma training corpus, the OLMo-core training code, every intermediate checkpoint, and the OLMES evaluation harness are all public. That's rare even among nominally 'open' models, and it makes OLMo the obvious pick for researchers who need to audit, reproduce or fine-tune from a known-clean base rather than trust a vendor's claims.
OLMo 3 closes the capability gap that dogged earlier releases — Ai2 says the 32B model beats other open releases at its scale, and OLMo 3.1 Think reportedly beats Qwen 3 32B on AIME 2025 — while training 2.5x more efficiently than Llama 3.1 per GPU-hour. It's still not a hosted product: no chat app, no API of its own. You run it yourself via Hugging Face, Ollama or vLLM. Free under Apache 2.0, usable anywhere including ASEAN for self-hosting.
What people say
OLMo's pitch has never been 'the smartest model' — it's 'the only model you can actually audit.' Where most open-weight releases from Meta, Mistral or Alibaba ship weights and call it open source, Ai2 (Allen Institute for AI) publishes the full stack: Dolma 2, the 7-trillion-token training corpus; OLMo-core, the actual training code; every intermediate checkpoint from pretraining; OLMES, the evaluation harness used to score it; and Tulu 3, the post-training recipe. Apache 2.0 licensing covers all of it.
That transparency used to come at a capability cost. OLMo 3, the current generation, narrows that gap substantially — Ai2 describes the 32B variant as the best open-source language model released to date at its scale, and says OLMo 3 trains 2.5x more efficiently than Meta's Llama 3.1 measured in GPU-hours per token. The interconnects.ai writeup, run by AI researcher Nathan Lambert who has direct visibility into the open-model space, called OLMo 3 'America's truly open reasoning models,' framing it against the wave of Chinese open releases like DeepSeek and Qwen as a U.S.-built alternative with real reasoning capability, not just instruction-following. A follow-up point release, OLMo 3.1, added extended reinforcement learning training; Ai2 reports OLMo 3.1 Think outperforming Qwen 3 32B on the AIME 2025 math benchmark and landing close to Gemma 27B.
Context length is a real limitation versus frontier proprietary models — support tops out around 65,000 tokens, roughly a short book chapter, well short of the 1M-token windows now common on Gemini or Grok. And OLMo isn't a product in the conventional sense: there's no Ai2-hosted chat interface or API tier to point at, no pricing page, no support contract. You pull weights from Hugging Face and run them yourself via Ollama, vLLM or the OLMo-core stack, which puts real infrastructure work between download and production use.
The audience this serves well is narrow but specific: researchers who need reproducibility, academics who need to cite exact training conditions, and organizations with license or provenance requirements that rule out murkier 'open-weight' releases. Ai2 CEO Ali Farhadi's own framing — that 'openness and performance can advance together' — is a fair summary of what OLMo 3 actually delivers.
Summary of public user & expert reviews, compiled by RECATOOLS.
About this listing
This entry was compiled from publicly available data including OLMo by AllenAI's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with OLMo by AllenAI unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to OLMo by AllenAI directly →
Spotted something out of date? Suggest an update →
OLMo by AllenAI in the news
Alternatives to OLMo by AllenAI
More in LLMs & Chat
