Cohere's open North Small Translate beats DeepL and Google Translate
A 25B-active MoE, CC BY-NC and free on Cohere's API, that self-hosts on two H100s. Also today: an open SAM 3 matting upgrade and California's AI-auditor laws.
Copy markdown
Open weights that top the incumbents
North Small Translate is a 218B-parameter mixture-of-experts model with only 25B active, and in Cohere's own WMT26 evals it scores 83.60 — ahead of DeepL NextGen at 81.37 and Google Translate at 68.20. On long-context translation the gap is stark: 48.9 versus Google's 21.3.
The catch: the weights are non-commercial
It ships under CC BY-NC 4.0 — free to research and self-host, but shipping it in a product means a commercial license through Cohere's sales team and Model Vault. If you'd rather not host, it's callable free on Cohere's API tier.
It fits on hardware you can actually rent
The sparse MoE fires 8 of 128 experts per token, with 16K-in / 16K-out context, in BF16, FP8 or NVFP4 — from a single B200 down to two H100s. Load it with Hugging Face transformers or vLLM.
Build this weekend: open matting that beats SAM 3
Feyn's MultiMatte, out today, is a promptable image-matting model: say what to keep and it alpha-mattes the rest, hair and fuzzy edges included. It's a 19.5M-parameter LoRA on SAM 3 that lifts DIS-VD S-measure to 0.901 from 0.667, and installs with pip install nobg.
Elsewhere: California writes the AI-audit rulebook
Gov. Newsom signed SB 813 and AB 1405 on Sept 9, standing up a state registry of independent AI auditors and a framework for verifying model compliance. If you ship models or hiring tools into California, third-party audits are becoming a real requirement.