#Olmo3
LightNews LightNews
natolambert.bsky.social

Paper: allenai.org/papers/olmo3
Artifacts: huggingface.co/collections/...
Demo: playground.allenai.org
Interconnects post: www.interconnects.ai/p/olmo-3-ame...
Technical Ai2 Blog: allenai.org/blog/olmo3
allenai.org
November 20, 2025 at 2:32 PM
natolambert.bsky.social
Olmo 3 is a banger paper. Finally getting to read the rest of it 😅.

We love releasing things that serve as a comprehensive snapshot of public knowledge on training leading language models.

There's an award for whoever finds all the secrets first in the new arxiv version
allenai.org/papers/olmo3
November 29, 2025 at 7:57 PM
natolambert.bsky.social
My favorite of the Olmo 3 paper: Transparent auditing of the cost for the v3 models (not 3.1), based of wall clock time for pre/post train, evals, cluster issues, etc, as a counter to the famous $5.576M for DeepSeek V3

at $2/H100 hour, Olmo 3 start to end would cost $2.75M
allenai.org/papers/olmo3
December 12, 2025 at 6:43 PM
infodocket.bsky.social
Allen #AI Announces Launch of Olmo 3 allenai.org/blog/olmo3 Demos: playground.allenai.org #opensource #LLMs @ai2.bsky.social
November 20, 2025 at 4:43 PM
weibac.bsky.social
having great fun making olmo3-7b-instruct hallucinate descriptions of SCP-055

it always picks up on the antimemetic aspect, but describes the object anyways
December 26, 2025 at 8:44 PM
davidcrespo.bsky.social
here's a 32B model with fully open training data and training process, not just open weights. 32B isn't very big but it's probably good enough to be useful for *something*. I'm curious whether the hard line applies to something like this by association
allenai.org/blog/olmo3
allenai.org/olmo
Olmo from Ai2
Our fully open language model and complete model flow.
allenai.org
January 8, 2026 at 11:44 PM
ai2.bsky.social
✨ Try Olmo 3 in the Ai2 Playground → playground.allenai.org?utm_source=b... & our Discord → discord.gg/ai2
💻 Download: huggingface.co/collections/...
📝 Blog: allenai.org/blog/olmo3?u...
📚 Technical report: allenai.org/papers/olmo3...
Join the Ai2 Discord Server!
The official Discord for Ai2 (The Allen Institute for AI), a nonprofit AI lab. | 2814 members
discord.gg
November 20, 2025 at 2:37 PM
realbreetaylor.bsky.social
Re: offline AI:
I’m new to this; I won’t officially endorse a specific build/stack, but yes that little machine is a fully functional open-source (NOT: open-“weight”) LLM.

5500xt-Vulkan API-Ollama (NOT: Llama = Meta/Zsuck)-Olmo3

Olmo-3 (Open Lang. MO-del) is associated to the HuggingFace breach.
September 19, 2026 at 2:34 PM
timkellogg.me
Olmo 3 7B & 32B base & thinking models

@ai2.bsky.social has done it again, fully open models, fully open process

seems competitive with Qwen 3, excel you can fully reproduce any part of the training process

allenai.org/blog/olmo3
Two-panel line chart on a dark teal background showing model performance during “Base Model Training” (left panel) and “Post-Training” (right panel). Axes and panels: • Left panel title: “Base Model Training.” Left y-axis label: “Base Eval Average (%),” running roughly 50–100. Text along the bottom: “includes Pretraining, Midtraining, Long Context.” • Right panel title: “Post-Training.” Right y-axis label: “Adapt Eval Average (%),” running roughly 30–80. Bottom text: “includes SFT, DPO, RL.” Main pink curve: A bright pink line with diamond or cross-shaped markers runs across both panels, climbing steadily. • In Base Model Training it starts a bit above 60%, rises to about 70%, stays level, then jumps to the mid-70s. A label “Marin 32B” sits near one mid-range point, with a small white sailboat icon on a nearby dashed line. • In Post-Training it continues upward, with three points in the high-70s to low-80s, forming a nearly flat top. Each point is marked with a pink outlined diamond plus icon. Other models / baselines: Several white or pink dashed lines with different icons represent comparison models. • In the left panel a white dashed line with a sailboat icon starts just above 50% and climbs toward the mid-60s, and another dashed line with a white square-plus “medical cross” icon stays in the high-50s. • In the right panel, a dashed pink line labeled “OLMo 2 32B” rises from the low-40s to just under 50%. A white dashed line labeled “Apertus 70B” hovers around 60% with a square-plus icon. Top-right labels: Near the upper-right corner of the Post-Training panel, clustered labels mark high-performing models: • “Qwen 3 32B” and “OLMo 3” at the very top near ~80% on the Adapt Eval axis, with star-like white markers. • Slightly below, “Gemma 3 27B” and “Qwen 2.5 32B” around the low-60s.
November 20, 2025 at 2:19 PM
stellali.bsky.social
Model-specific patterns reveal training impact:

Olmo3 exhibits more diverse cognitive elements (49%)—they explicitly included meta-reasoning data during midtraining.
DeepHermes-3: only 12% avg presence.

Training methodology shapes cognitive profiles dramatically.
November 25, 2025 at 6:26 PM
natolambert.bsky.social
New Olmo 3.1 artifacts: huggingface.co/collections/...
Paper (arxiv soon): allenai.org/papers/olmo3
Demo: playground.allenai.org
Olmo 3.1 - a allenai Collection
The latest members of the Olmo 3 family: another 3 weeks of RL for 32B Think, the 32B Instruct model, large post-training research datasets...
huggingface.co
December 12, 2025 at 5:15 PM
timkellogg.me
the key: bootstrap from OLMo3
In the first stage, we freeze the Olmo 3 transformer and train only the local encoder, local decoder, boundary predictor, and language modeling head. This stage is designed to be cheap and fast, requiring only 9.8B tokens (approximately 43B bytes) to learn to mimic the subword model's behavior. In the second stage, we unfreeze the entire model and train for 39.3B additional tokens (approximately 173B bytes), letting Bolmo fully exploit byte-level information while maintaining efficiency.
December 16, 2025 at 3:27 PM
philpax.me
if you take what they're saying at face value, it's possible that they're still training (including on real data) and they're not releasing until they're happy there's no more juice to squeeze

similar thing happened with Olmo where they just kept cooking for a few more weeks allenai.org/blog/olmo3
March 23, 2026 at 3:06 PM
kylelo.bsky.social
🍏Try the model: playground.allenai.org
🍍Download the collection: huggingface.co/collections/...
🍌Read the blog: allenai.org/blog/olmo3
🍎And our 100+ page paper lol www.datocms-assets.com/64837/176364...
playground.allenai.org
November 20, 2025 at 6:20 PM
kylelo.bsky.social
new olmo 3.1 artifacts: huggingface.co/collections/...
paper (arxiv soon): allenai.org/papers/olmo3
demo: playground.allenai.org
Olmo 3.1 - a allenai Collection
The latest members of the Olmo 3 family: another 3 weeks of RL for 32B Think, the 32B Instruct model, large post-training research datasets...
huggingface.co
December 12, 2025 at 6:03 PM
yetanotheruseless.com
thus you should run your own AI, trained by an nonprofit who tells you the exact recipe of what when into it, and tells you how to modify your own variation of it (e.g. allenai.org/blog/olmo3 ).

Schools and governments should *definitely* follow this route.
Olmo 3: Charting a path through the model flow to lead open-source AI | Ai2
Our new flagship Olmo 3 model family empowers the open source community with not only state-of-the-art open models, but the entire model flow and full traceability back to training data.
allenai.org
November 23, 2025 at 11:41 PM
comentadortortuga.bsky.social
toma, el código de un modelo Olmo, lo puedes ver github.com/allenai/OLMo...
OLMo-core/src/scripts/official/OLMo3/README.md at main · allenai/OLMo-core
PyTorch building blocks for the OLMo ecosystem. Contribute to allenai/OLMo-core development by creating an account on GitHub.
github.com
September 14, 2026 at 10:10 PM
timkellogg.me
on #3, this paper uses a method where they can directly attribute specific documents from the pretraining dataset

they used it to show that LLMs do in fact learn procedures, not just autocomplete. But you could take this so much further with Olmo3

arxiv.org/abs/2411.12580
A comic-style infographic titled “THE AI CHEF’S ‘PROCEDURAL’ SECRET: AN ATTRIBUTION ANALOGY.” It uses a robot chef baking a soufflé to explain how attribution and gradient-based tracing in AI works. The diagram proceeds left to right in five labeled steps. ⸻ 1. THE TASK (REASONING) A friendly robot chef stands in a kitchen, holding up a perfectly baked soufflé. A math bubble shows x + 2y = 10 as an analogy for solving a problem. Caption: AI Chef (LLM) solves a problem (bakes a soufflé). ⸻ 2. THE “FINGERPRINT” (GRADIENT) Close-up of the robot whisking batter. A glowing network of abstract swirls appears over the bowl. Caption: We record the exact, unique actions & “effort” (Gradient) used for this specific soufflé. ⸻ 3. THE “BRAIN MAP” (EK/FAC) The robot stands before floating diagram bubbles labeled Whisking Techniques, Aeration Physics, Heat Transfer, Simplified Linkages. Caption: We use a simplified map of how the chef connects concepts (Hessian/EK-FAC approximation). ⸻ 4. THE LIBRARY MATCH (ATTRIBUTION) The robot enters a vast library with floor-to-ceiling bookshelves. A giant glowing fingerprint projection shines onto one shelf as the robot scans for the best match. Caption: We scan the entire “cookbook library” (pre-training data) to find which book’s instructions best match the fingerprint via the brain map. ⸻ 5. THE RESULT: PROCEDURAL KNOWLEDGE The robot chef proudly holds a glowing lightbulb while a book opens nearby with a concept diagram. A large reference book beside him is titled “THE PHYSICS OF FOAMS & AERATION (NOT a Soufflé Recipe Book!)” Caption: We find the source was NOT a recipe, but a foundational PRINCIPLE (procedural knowledge) applied to a new task. ⸻ Overall, the image uses the story of baking a soufflé to explain how AI models trace reasoning: capturing gradients, mapping conceptual relations, searching training data, and revealing underlying procedural knowledge rather than direct memorization.
November 23, 2025 at 3:20 PM
osma.sigmoid.social.ap.brid.gy
[math & AI dadjoke]

The Public: AI is the root of all evil!

Allen Institute (Ai2): We're not evil, we just released a new version of the world's most open LLM: https://allenai.org/blog/olmo3
奥尔莫3：开辟模型流路径，引领开源人工智能
Olmo 3: Charting a path through the model flow to lead open-source AI (allenai.org) 14:50  ↑ 105 HN Points
allenai.org
November 23, 2025 at 11:47 AM
verdverm.com
allenai.org/blog/olmo3
Olmo 3: Charting a path through the model flow to lead open-source AI | Ai2
Our new flagship Olmo 3 model family empowers the open source community with not only state-of-the-art open models, but the entire model flow and full traceability back to training data.
allenai.org
November 21, 2025 at 6:55 AM
winbuzzer.com
winbuzzer.com/2025/11/21/a...

AI2 Releases OLMo 3: A Fully Open 'Model Flow' to Challenge Black Box AI Paradigm

#AI #OpenSource #OLMo3 #LLM #AI2 #GenAI #MachineLearning #AIResearch #OpenScience #DeepLearning #AIModels #EthicalAI
AI2 Releases OLMo 3: A Fully Open 'Model Flow' to Challenge Black Box AI Paradigm - WinBuzzer
The Allen Institute for AI has launched OLMo 3, releasing the complete "model flow", including datasets and training code, to challenge the industry's black-box AI standards.
winbuzzer.com
November 21, 2025 at 8:43 AM
techlife-blog.bsky.social
Olmo 3 Revolutionizes Open-Source AI with Transparent Model Flow
You must try !!!
techlife.blog/posts/olmo-3...

#OpenSource #LanguageModels #AI #OpensourceAI #Olmo3
Olmo 3 Revolutionizes Open-Source AI with Transparent Model Flow
Olmo 3 introduces a groundbreaking approach to open-source AI development by making the entire model flow accessible and customizable.
techlife.blog
November 23, 2025 at 11:50 AM
sfresearch.bsky.social
(4/8) Against the strongest self-distillation baseline (SDPO):

Qwen3-8B 📐

SciKnowEval: 74.4 → 77.2

💻 LiveCodeBench: 47.9 → 51.7

OLMo3-7B 📐

Science: 69.5 → 73.3

💻 Code: 45.0 → 51.1

Same backbone. Only difference: memory.
June 3, 2026 at 10:03 PM
handle.invalid
Olmo 3: Charting a path through the model flow to lead open-source AI

https://allenai.org/blog/olmo3

#machinelearning #datascience
Olmo 3: Charting a path through the model flow to lead open-source AI
Summary: Olmo 3 unveils open model flow with traceable data, 7B/32B models, Think/Instruct/RL Zero paths, and Dolma 3 data curriculum.
allenai.org
November 22, 2025 at 12:12 AM
geodesicresearch.bsky.social
However, we found alignment pretraining was even more useful than expected, and improved the behaviours of our post-trained models (trained on over 4.5M examples of desired AI behaviour from @natolambert.bsky.social's and Olmo3's full post-training mix).
May 19, 2026 at 7:31 AM