Two big LLM releases today - Cohere's Command A simonwillison.net/2025/Mar/13/... and Ai2's OLMo 2
OLMo claims to be "the first fully-open model (all data, code, weights, and details are freely available) to outperform GPT3.5-Turbo and GPT-4o mini", which feels notable allenai.org/blog/olmo2-32B
OLMo claims to be "the first fully-open model (all data, code, weights, and details are freely available) to outperform GPT3.5-Turbo and GPT-4o mini", which feels notable allenai.org/blog/olmo2-32B
Introducing Command A: Max performance, minimal compute
New LLM release from Cohere. It's interesting to see which aspects of the model they're highlighting, as an indicator of what their commercial customers value the most (highlight mine): > …
simonwillison.net
March 13, 2025 at 9:32 PM
Two big LLM releases today - Cohere's Command A simonwillison.net/2025/Mar/13/... and Ai2's OLMo 2
OLMo claims to be "the first fully-open model (all data, code, weights, and details are freely available) to outperform GPT3.5-Turbo and GPT-4o mini", which feels notable allenai.org/blog/olmo2-32B
OLMo claims to be "the first fully-open model (all data, code, weights, and details are freely available) to outperform GPT3.5-Turbo and GPT-4o mini", which feels notable allenai.org/blog/olmo2-32B
Yes.
Good to see work like Olmo2 and Nous DisTrO too. Momentum is building.
allenai.org/blog/olmo2
bsky.app/profile/nous...
Good to see work like Olmo2 and Nous DisTrO too. Momentum is building.
allenai.org/blog/olmo2
bsky.app/profile/nous...
December 3, 2024 at 1:29 AM
Yes.
Good to see work like Olmo2 and Nous DisTrO too. Momentum is building.
allenai.org/blog/olmo2
bsky.app/profile/nous...
Good to see work like Olmo2 and Nous DisTrO too. Momentum is building.
allenai.org/blog/olmo2
bsky.app/profile/nous...
What a crazy few weeks launching multiple state-of-the-art AI models at Ai2. Tulu 3 post training, OLMo2, Molmo just in the last month.
I've learned a lot about what makes good model training teams work and written it up to celebrate these releases. All about focus.
https://buff.ly/3COWzs7
I've learned a lot about what makes good model training teams work and written it up to celebrate these releases. All about focus.
https://buff.ly/3COWzs7
November 27, 2024 at 4:00 PM
What a crazy few weeks launching multiple state-of-the-art AI models at Ai2. Tulu 3 post training, OLMo2, Molmo just in the last month.
I've learned a lot about what makes good model training teams work and written it up to celebrate these releases. All about focus.
https://buff.ly/3COWzs7
I've learned a lot about what makes good model training teams work and written it up to celebrate these releases. All about focus.
https://buff.ly/3COWzs7
Demo: playground.allenai.org
Blog: allenai.org/blog/olmo2
Data & Weights: huggingface.co/collections/...
Ai2 post: bsky.app/profile/ai2....
finishing the tech report, can't wait to share it soon!
Blog: allenai.org/blog/olmo2
Data & Weights: huggingface.co/collections/...
Ai2 post: bsky.app/profile/ai2....
finishing the tech report, can't wait to share it soon!
playground.allenai.org
November 26, 2024 at 8:59 PM
Demo: playground.allenai.org
Blog: allenai.org/blog/olmo2
Data & Weights: huggingface.co/collections/...
Ai2 post: bsky.app/profile/ai2....
finishing the tech report, can't wait to share it soon!
Blog: allenai.org/blog/olmo2
Data & Weights: huggingface.co/collections/...
Ai2 post: bsky.app/profile/ai2....
finishing the tech report, can't wait to share it soon!
Results: models actively use cross-layer retrieval, the attention-sink phenomenon disappears, and MoDA improves the OLMo2 baseline across the board.
📄 Paper: arxiv.org/abs/2603.15619
✍️ Blog: lh-zhu.github.io/The-Second-H...
💻 Code: github.com/hustvl/MoDA
📄 Paper: arxiv.org/abs/2603.15619
✍️ Blog: lh-zhu.github.io/The-Second-H...
💻 Code: github.com/hustvl/MoDA
April 20, 2026 at 1:00 AM
Results: models actively use cross-layer retrieval, the attention-sink phenomenon disappears, and MoDA improves the OLMo2 baseline across the board.
📄 Paper: arxiv.org/abs/2603.15619
✍️ Blog: lh-zhu.github.io/The-Second-H...
💻 Code: github.com/hustvl/MoDA
📄 Paper: arxiv.org/abs/2603.15619
✍️ Blog: lh-zhu.github.io/The-Second-H...
💻 Code: github.com/hustvl/MoDA
#OLMo2 32B: First fully #opensource model to outperform #GPT3.5 and #GPT4o mini 🔥
🧵👇 #MachineLearning #AI #llm
🧵👇 #MachineLearning #AI #llm
March 16, 2025 at 7:14 AM
#OLMo2 32B: First fully #opensource model to outperform #GPT3.5 and #GPT4o mini 🔥
🧵👇 #MachineLearning #AI #llm
🧵👇 #MachineLearning #AI #llm
After #Teuken7b and #Olmo2 , Apertus is the next big jump in capabilities and performance of #FOSS #LLMs , while also improving #epistemicresilience and #epistemicautonomy with its multilingual approach. [4/5]
September 2, 2025 at 8:34 AM
After #Teuken7b and #Olmo2 , Apertus is the next big jump in capabilities and performance of #FOSS #LLMs , while also improving #epistemicresilience and #epistemicautonomy with its multilingual approach. [4/5]
Read about the OLMo 2 recipe in our blog: allenai.org/blog/olmo2
Download the full OLMo 2 collection, including model weights and data on HuggingFace: huggingface.co/collections/...
Access the training code on GitHub: github.com/allenai/OLMo
Download the full OLMo 2 collection, including model weights and data on HuggingFace: huggingface.co/collections/...
Access the training code on GitHub: github.com/allenai/OLMo
November 26, 2024 at 8:53 PM
Read about the OLMo 2 recipe in our blog: allenai.org/blog/olmo2
Download the full OLMo 2 collection, including model weights and data on HuggingFace: huggingface.co/collections/...
Access the training code on GitHub: github.com/allenai/OLMo
Download the full OLMo 2 collection, including model weights and data on HuggingFace: huggingface.co/collections/...
Access the training code on GitHub: github.com/allenai/OLMo
Paper: arxiv.org/abs/2501.00656
Blog: allenai.org/blog/olmo2
Demo: playground.allenai.org
Collection: huggingface.co/collections/...
Blog: allenai.org/blog/olmo2
Demo: playground.allenai.org
Collection: huggingface.co/collections/...
2 OLMo 2 Furious
We present OLMo 2, the next generation of our fully open language models. OLMo 2 includes dense autoregressive models with improved architecture and training recipe, pretraining data mixtures, and ins...
arxiv.org
January 6, 2025 at 8:28 PM
Paper: arxiv.org/abs/2501.00656
Blog: allenai.org/blog/olmo2
Demo: playground.allenai.org
Collection: huggingface.co/collections/...
Blog: allenai.org/blog/olmo2
Demo: playground.allenai.org
Collection: huggingface.co/collections/...
🔗 Blog: https://allenai.org/blog/olmo2
🔗 HF Collection: https://huggingface.co/collections/allenai/olmo-2-674117b93ab84e98afc72edc
🔗 HF Collection: https://huggingface.co/collections/allenai/olmo-2-674117b93ab84e98afc72edc
OLMo 2 - a allenai Collection
huggingface.co
November 26, 2024 at 9:37 PM
OLMo doesn’t match 4, but it does come fairly close. Your question remains legit though. allenai.org/blog/olmo2-32b
OLMo 2 32B: First fully open model to outperform GPT 3.5 and GPT 4o mini | Ai2
Introducing OLMo 2 32B, the most capable and largest model in the OLMo 2 family.
allenai.org
August 8, 2025 at 12:57 PM
OLMo doesn’t match 4, but it does come fairly close. Your question remains legit though. allenai.org/blog/olmo2-32b
"Supposedly Equivalent Facts That Aren't? Entity Frequency in Pre-training Induces Asymmetry in LLMs" by Yuan He et al. arxiv.org/abs/2503.22362
October 14, 2025 at 6:19 PM
"Supposedly Equivalent Facts That Aren't? Entity Frequency in Pre-training Induces Asymmetry in LLMs" by Yuan He et al. arxiv.org/abs/2503.22362
We have the answers of these questions here : arxiv.org/pdf/2509.22367
We analyze the political content of the training data from OLMO2, the largest fully open-source model.
🕵️♀️ We run an analysis in all the datasets (2 pre- and 2 post-training) used to train the models. Here are our findings:
We analyze the political content of the training data from OLMO2, the largest fully open-source model.
🕵️♀️ We run an analysis in all the datasets (2 pre- and 2 post-training) used to train the models. Here are our findings:
arxiv.org
September 29, 2025 at 2:54 PM
We have the answers of these questions here : arxiv.org/pdf/2509.22367
We analyze the political content of the training data from OLMO2, the largest fully open-source model.
🕵️♀️ We run an analysis in all the datasets (2 pre- and 2 post-training) used to train the models. Here are our findings:
We analyze the political content of the training data from OLMO2, the largest fully open-source model.
🕵️♀️ We run an analysis in all the datasets (2 pre- and 2 post-training) used to train the models. Here are our findings:
I think I know what you mean, I've seen it make factual errors as well; have you had a chance to eval Tulu3 vs Olmo2?
December 3, 2024 at 11:06 PM
I think I know what you mean, I've seen it make factual errors as well; have you had a chance to eval Tulu3 vs Olmo2?
8/9 Furthermore, models that show a small diff in perplexity can have a large number of features where they differ!
Comparisons between Llama2 and OLMo2 models of the same size (which barely show a diff in perplexity) had the greatest number of discovered features.
Comparisons between Llama2 and OLMo2 models of the same size (which barely show a diff in perplexity) had the greatest number of discovered features.
June 9, 2025 at 1:47 PM
8/9 Furthermore, models that show a small diff in perplexity can have a large number of features where they differ!
Comparisons between Llama2 and OLMo2 models of the same size (which barely show a diff in perplexity) had the greatest number of discovered features.
Comparisons between Llama2 and OLMo2 models of the same size (which barely show a diff in perplexity) had the greatest number of discovered features.
For OLMo 2, we focused on 4 things:
Training stability
Better data curriculum in pretraining
State of the art post-training recipes
Better evals to help us make decisions as we develop
Read more below 🧵 or in our blogpost: allenai.org/blog/olmo2
Training stability
Better data curriculum in pretraining
State of the art post-training recipes
Better evals to help us make decisions as we develop
Read more below 🧵 or in our blogpost: allenai.org/blog/olmo2
OLMo 2: The best fully open language model to date | Ai2
Our next generation of fully-open base and instruct models sit at the Pareto frontier of performance and training efficiency.
allenai.org
November 26, 2024 at 9:12 PM
For OLMo 2, we focused on 4 things:
Training stability
Better data curriculum in pretraining
State of the art post-training recipes
Better evals to help us make decisions as we develop
Read more below 🧵 or in our blogpost: allenai.org/blog/olmo2
Training stability
Better data curriculum in pretraining
State of the art post-training recipes
Better evals to help us make decisions as we develop
Read more below 🧵 or in our blogpost: allenai.org/blog/olmo2
The release that I am excited about:
allenai.org/blog/olmo2?u...
allenai.org/blog/olmo2?u...
OLMo 2: The best fully open language model to date | Ai2
Our next generation of fully-open base and instruct models sit at the Pareto frontier of performance and training efficiency.
allenai.org
December 5, 2024 at 6:27 PM
The release that I am excited about:
allenai.org/blog/olmo2?u...
allenai.org/blog/olmo2?u...
Text-generation-inference v3.0.2 is out.
Basically we can run transformers models (that support flash) at roughly the same speed as native TGI ones.
What this means is broader model support.
Today it unlocks
Cohere2, Olmo, Olmo2 and Helium
Congrats Cyril Vallez
github.com/huggingface/...
Basically we can run transformers models (that support flash) at roughly the same speed as native TGI ones.
What this means is broader model support.
Today it unlocks
Cohere2, Olmo, Olmo2 and Helium
Congrats Cyril Vallez
github.com/huggingface/...
Release v3.0.2 · huggingface/text-generation-inference
Tl;dr
New transformers backend supporting flashattention at roughly same performance as pure TGI for all non officially supported models directly in TGI. Congrats @Cyrilvallez
New models unlocked: ...
github.com
January 24, 2025 at 2:55 PM
Text-generation-inference v3.0.2 is out.
Basically we can run transformers models (that support flash) at roughly the same speed as native TGI ones.
What this means is broader model support.
Today it unlocks
Cohere2, Olmo, Olmo2 and Helium
Congrats Cyril Vallez
github.com/huggingface/...
Basically we can run transformers models (that support flash) at roughly the same speed as native TGI ones.
What this means is broader model support.
Today it unlocks
Cohere2, Olmo, Olmo2 and Helium
Congrats Cyril Vallez
github.com/huggingface/...
6/ Multilinguality shapes layerwise pivoting, most clearly in the decoding probe. Aya23 and Apertus show other-language pivoting from earlier layers than Llama2 and OLMo2. In the representation-based probe, the pattern is less clear-cut, with OLMo2 standing out as an exception.
September 8, 2026 at 12:21 PM
6/ Multilinguality shapes layerwise pivoting, most clearly in the decoding probe. Aya23 and Apertus show other-language pivoting from earlier layers than Llama2 and OLMo2. In the representation-based probe, the pattern is less clear-cut, with OLMo2 standing out as an exception.
✅ Supports regional innovation and infrastructure
✅ Advances regulatory and technological sovereignty 🛠 From small models like OLMo2 to tools like Hugging Face Transformers or Sarvam-M for Indian languages, OS efforts are already powering sovereign AI ecosystems worldwide.
✅ Advances regulatory and technological sovereignty 🛠 From small models like OLMo2 to tools like Hugging Face Transformers or Sarvam-M for Indian languages, OS efforts are already powering sovereign AI ecosystems worldwide.
June 11, 2025 at 3:13 PM
✅ Supports regional innovation and infrastructure
✅ Advances regulatory and technological sovereignty 🛠 From small models like OLMo2 to tools like Hugging Face Transformers or Sarvam-M for Indian languages, OS efforts are already powering sovereign AI ecosystems worldwide.
✅ Advances regulatory and technological sovereignty 🛠 From small models like OLMo2 to tools like Hugging Face Transformers or Sarvam-M for Indian languages, OS efforts are already powering sovereign AI ecosystems worldwide.
For more information, read the blog post: allenai.org/blog/olmo2-32B
Or try the model on the Ai2 Playground: playground.allenai.org?model=olmo-2...
Get the artifacts here: huggingface.co/collections/...
Or try the model on the Ai2 Playground: playground.allenai.org?model=olmo-2...
Get the artifacts here: huggingface.co/collections/...
OLMo 2 32B: First fully open model to outperform GPT 3.5 and GPT 4o mini | Ai2
Introducing OLMo 2 32B, the most capable and largest model in the OLMo 2 family.
allenai.org
March 13, 2025 at 6:36 PM
For more information, read the blog post: allenai.org/blog/olmo2-32B
Or try the model on the Ai2 Playground: playground.allenai.org?model=olmo-2...
Get the artifacts here: huggingface.co/collections/...
Or try the model on the Ai2 Playground: playground.allenai.org?model=olmo-2...
Get the artifacts here: huggingface.co/collections/...
So as I was saying yesterday ;)
allenai.org/blog/olmo2-32B
Huge congrats to the @ai2.bsky.social team. This is a fantastic achievement, and a strong reminder not to discount meaningfully open models when talking about the state of the art in AI!!!
allenai.org/blog/olmo2-32B
Huge congrats to the @ai2.bsky.social team. This is a fantastic achievement, and a strong reminder not to discount meaningfully open models when talking about the state of the art in AI!!!
March 13, 2025 at 6:47 PM
So as I was saying yesterday ;)
allenai.org/blog/olmo2-32B
Huge congrats to the @ai2.bsky.social team. This is a fantastic achievement, and a strong reminder not to discount meaningfully open models when talking about the state of the art in AI!!!
allenai.org/blog/olmo2-32B
Huge congrats to the @ai2.bsky.social team. This is a fantastic achievement, and a strong reminder not to discount meaningfully open models when talking about the state of the art in AI!!!