Today we’re announcing *Olmo 3*—our leading fully open language model suite built for reasoning, chat, and tool use, & an open model flow that exposes not just the final weights, but the entire training journey.
Most models ship as a single opaque snapshot. Olmo 3 opens the model flow end to end – pretraining, mid-training, and post-training – plus data recipes and code, so you can see how capabilities are built and customize any stage of the process.
Meet the Olmo 3 family:
*Olmo 3-Base (7B, 32B)*—foundations for post-training with strong code, math, and reading comprehension skills
*Olmo 3-Instruct (7B)*—focused on multi-turn chat and tool use
*Olmo 3-Think (7B, 32B)*—“thinking” models that surface their reasoning steps
All are compact, dense models designed to run on hardware ranging from laptops to research clusters.
Under the hood, we trained Olmo 3 on ~6T tokens from our new *Dolma 3* pretraining dataset, plus new post-training sets with stronger data decontamination and richer math/code/reasoning mixes. A long-context extension pushes Olmo 3’s context window to ~65K tokens—enough for full papers, books, and other long files.
At the center is *Olmo 3-Think (32B)*, the best fully open 32B-scale reasoning model we’re aware of, alongside our strongest 32B base model.
*In our evaluations:*
⦿ Olmo 3-Think (32B) is the *strongest fully open 32B-scale reasoning model*
⦿ Olmo 3-Base models *beat fully open Marin & Apertus *and* rival Qwen 2.5 and Gemma 3*
⦿ Olmo 3-Instruct (7B) *beats Qwen 2.5, Gemma 3, and Llama 3.1 on tough chat + tool-use benchmarks*
We’re also rolling out a major Ai2 Playground upgrade alongside Olmo 3:
*Thinking mode* to see intermediate reasoning on complex tasks
*Tool calling* so you can define JSON-schema tools or call tools via our Asta platform
Olmo 3 is wired into *OlmoTrace* in the Ai2 Playground, so you don’t just see its behavior—you can trace it. For example, you can ask Olmo 3-Think (32B) to answer a general-knowledge question, then use OlmoTrace to inspect where and how the model may have learned to generate parts of its response.
If you care about AI you can customize, inspect, and improve, Olmo 3 is for you—available now under Apache 2.0.
Today we’re announcing *Olmo 3*—our leading fully open language model suite built for reasoning, chat, and tool use, & an open model flow that exposes not just the final weights, but the entire training journey.
Most models ship as a single opaque snapshot. Olmo 3 opens the model flow end to end – pretraining, mid-training, and post-training – plus data recipes and code, so you can see how capabilities are built and customize any stage of the process.
Meet the Olmo 3 family:
*Olmo 3-Base (7B, 32B)*—foundations for post-training with strong code, math, and reading comprehension skills
*Olmo 3-Instruct (7B)*—focused on multi-turn chat and tool use
*Olmo 3-Think (7B, 32B)*—“thinking” models that surface their reasoning steps
All are compact, dense models designed to run on hardware ranging from laptops to research clusters.
Under the hood, we trained Olmo 3 on ~6T tokens from our new *Dolma 3* pretraining dataset, plus new post-training sets with stronger data decontamination and richer math/code/reasoning mixes. A long-context extension pushes Olmo 3’s context window to ~65K tokens—enough for full papers, books, and other long files.
At the center is *Olmo 3-Think (32B)*, the best fully open 32B-scale reasoning model we’re aware of, alongside our strongest 32B base model.
*In our evaluations:*
⦿ Olmo 3-Think (32B) is the *strongest fully open 32B-scale reasoning model*
⦿ Olmo 3-Base models *beat fully open Marin & Apertus *and* rival Qwen 2.5 and Gemma 3*
⦿ Olmo 3-Instruct (7B) *beats Qwen 2.5, Gemma 3, and Llama 3.1 on tough chat + tool-use benchmarks*
We’re also rolling out a major Ai2 Playground upgrade alongside Olmo 3:
*Thinking mode* to see intermediate reasoning on complex tasks
*Tool calling* so you can define JSON-schema tools or call tools via our Asta platform
Olmo 3 is wired into *OlmoTrace* in the Ai2 Playground, so you don’t just see its behavior—you can trace it. For example, you can ask Olmo 3-Think (32B) to answer a general-knowledge question, then use OlmoTrace to inspect where and how the model may have learned to generate parts of its response.
If you care about AI you can customize, inspect, and improve, Olmo 3 is for you—available now under Apache 2.0.
Dive deeper & get started:
Give Olmo 3 a spin in the Ai2 Playground → https://playground.allenai.org?utm_source=discord&utm_medium...
Download the models: https://huggingface.co/collections/allenai/olmo-3-68e80f043c...
Read more in our blog: https://allenai.org/blog/olmo3?utm_source=discord&utm_medium...
Check out the tech report: https://allenai.org/papers/olmo3?utm_source=discord&utm_medi...