Soofi S releases a 30B-A3B open Mamba-MoE model trained on 27T tokens
_ScottCondron · x · 2026-07-23
The post announces the Soofi S pretraining tech report, describing a 30B-A3B mixture-of-experts hybrid Mamba model trained on about 27T tokens with intentionally up-weighted German.
Key claims from the report:
- It is the strongest fully open model in the authors’ evaluations on both the English and German aggregates, ahead of Olmo 3 32B and Apertus 70B.
- The release includes unusually complete transparency: per-source data accounting, hyperparameters, training and eval code, checkpoints, and a project page.
- The project was trained on Deutsche Telekom infrastructure, and the authors even released a Weights & Biases dashboard for the run.
The post is basically celebrating how unusually open the whole training stack is, not just the model weights.
Related event: Soofi S Releases 30B Hybrid Mamba Open-Source Model(2 posts)→
More from Models
- Engram's random reads don't suit SSDs; CPU-memory over NVLink could serve all 72 GPUs — bookwormengr · 2026-09-11
- Meta's Muse Agent has built-in invite code logic, hinting at free-usage expansion — testingcatalog · 2026-09-11
- Same Echo Maze prompt, three frontier models: all passed visually but shipped the same hidden bug — eyishazyer · 2026-09-11
- Benchmark scores drop from 89% to 19% on new evals — how benchmaxxing breaks leaderboard trust — airesearch12 · 2026-09-11
- ChatGPT tells user their question is too hard and to 'accept dumber answers' — phido3000 · 2026-09-11
- Claude is no longer available for minors as Anthropic rolls out age assurance — Muhammad523 · 2026-09-11