Ai2 says Marin’s latest hero run includes about 1.8T Dolma 3.5 tokens

AAi29/25/2026, 14:20:03

Ai2 says Marin’s latest hero run includes about 1.8 trillion tokens from Ai2’s Dolma 3.5 corpus. Ai2 also says Marin’s data-mixture work draws on its Olmix recipes and the Organize the Web data-curation methodology.

Will Held says Marin’s 535B run was built on 25 trillion tokens from 152 datasets with licenses that permit training.

#marin #dolma #training-data #ai2 #olmix

View original post ↗

Previous articleJev-Mem paper reports faster memory construction for AI agents Next articleDSPy 3.4.0 adds native Jev and System One support

Related AI news

Ai2 makes Stage 1 byte-level checkpoints available with unchanged base weights

Ai2 has released Stage 1 checkpoints for its byte-level language models, keeping original model weights unchanged while providing pre-trained byte components. The release allows researchers to experiment with new architectures and train systems without re-running the initial stage.

AAi2•10/7/2026, 09:01:28•

#ai2 #byte-level-models #model-checkpoints #open-source-ai

Ai2 byte-level language model paper accepted to Nature, releases Qwen and Llama checkpoints

Ai2 announced that its paper on retrofitting language models to operate over bytes has been accepted to Nature, alongside releasing new checkpoints that extend the Bolmo method to Qwen and Llama.

AAi2•10/7/2026, 09:00:53•

#ai2 #bolmo #nature #language-models

Google Cloud matches Ai2’s Olmo 3 7B run on Cloud TPUs

Ai2 says Google Cloud reran Olmo 3’s 7B pretraining and mid-training on Cloud TPUs and matched the original run on held-out evaluations.

AAi2•9/29/2026, 08:58:33•

#olmo-3 #ai2 #google-cloud #tpu #reproducibility

Ai2 and GlobalFishWatch announce an AI partnership for ocean monitoring and enforcement

At Climate Week NYC, Ai2 and GlobalFishWatch announced a partnership on how AI and AI agents are used in ocean monitoring and enforcement. Ai2 describes the effort as purpose-built, open, and accountable to the people who use it.

AAi2•9/23/2026, 11:40:35•

#ai2 #globalfishwatch #ai-agents #ocean-monitoring

Marin 535B-A23B reaches training halfway mark as kernel updates lift MFU to 27%

Percy Liang and the Marin Community team have reached the halfway mark in training the open Marin 535B-A23B model on CoreWeave. Better kernels have lifted MFU from 21% to 27%, with heldout loss tracking forecasts built from four preliminary runs.

WWeights & Biases•10/7/2026, 11:38:52•

#marin #open-source-ai #weights-&-biases #coreweave

Reset history AI News FAQ Usage limits RSS feeds API MCP Skill

News is compiled from public information. The data is for reference only, and this site has no affiliation with OpenAI.

Data & original project: Did Codex Reset · Licoy