Cohere Releases North Small Translate: A 218B MoE Translation Model That Scores 83.6 on WMT26 Across 50 Languages

Details

Cohere released North Small Translate, the first dedicated translation model in its North family, built in partnership with RWS (Language Weaver). The architecture is a decoder-only sparse MoE Transformer: 218B total parameters, 25B active per forward pass, 128 experts with 8 activated per token plus a shared expert applied to every token. Attention alternates 3:1 between sliding-window layers (4K window, RoPE) and global layers, a layout first introduced in Command A. Context length is 16K tokens in and 16K out, text only. On Cohere's WMT26 evaluation — using GPT-5.6-Sol as the judge — the model scores 83.6 averaged across 50 languages, versus DeepL NextGen at 81.37, Qwen 3.5 397B at 81.56, and Google Translate at 68.20. An agentic variant with multi-pass self-correction reaches 84.36. These are vendor-reported numbers with a vendor-chosen judge; treat them as directional until independent WMT26 results appear. Three deployment paths: free via Cohere's API within rate limits, non-commercial self-hosting, or a commercial license. The model is open-weight but not fully open-source — a custom license governs commercial use. For teams running multilingual pipelines, North Small Translate offers a credible alternative to commercial translation APIs, particularly on languages where Google Translate underperforms. The 25B active parameter footprint keeps per-token compute tractable, though the full 218B weight load requires substantial VRAM for self-hosting.