Cohere Releases North Small Translate, a 218B MoE Translation Model

According to the report, the model can be called free on Cohere's API until rate limits apply, self-hosted non-commercially, or licensed commercially. Cohere built it with RWS, whose Language Weaver scientists and language experts shaped its real-world quality. North Small Translate is the first translation model in Cohere's North family. It follows Tiny Aya and Command A Translate in Cohere's multilingual lineage. In a post on X, Cohere framed translation as a sovereignty issue, saying organizations that cannot communicate globally cannot stay sovereign.

The model is a decoder-only sparse MoE Transformer. It uses 128 experts, with 8 activated per token and shared experts applied to every token. Its router uses a sigmoid over expert logits, normalized over the selected top-k. Attention alternates sliding-window layers with a window of 4096 and RoPE and global layers without positional embeddings, interleaved 3:1. That attention layout was first introduced in Command A. The context window is 16K input and 16K output tokens, text only. Cohere post-trained the model specifically for translation quality. About 11.5% of the weights are active per token; per-token compute tracks the 25 billion active parameters, while memory still has to hold all 218 billion. The release returns to the original translation problem that produced the Transformer's first results in 2017, when Google researchers introduced the architecture with Attention Is All You Need and evaluated it on WMT 2014 English-to-German and English-to-French.

Cohere's team reported these WMT26 all-languages scores in its launch blog: North Small Translate in agentic mode 84.36, North Small Translate 83.60, Qwen 3.5 397B A17B 81.56, DeepL NextGen 81.37, Gemma 4 31B (on) 79.46, GLM 5.2 FP8 76.50, and Google Translate 68.20. The agentic variant runs a multi-pass workflow that finds and fixes its own errors. Cohere's scoring bands treat 80 to 100 as perfect or minor errors only.

The scores come with a caveat. They are Cohere's own runs, with GPT-5.6-Sol as the judge, and should be treated as vendor-reported until independent WMT26 results appear. Regionally, both versions beat Gemma 4 31B (on) across Europe. On EU languages, the standard model scores 82.17 against Gemma's 72.73. South Asia is close, at 86.16 for North against 88.04 for Gemma.

In Cohere's tests, the model produced 112 output tokens per second against 81 for Gemma 4 31B at low concurrency on identical hardware. At high concurrency, the figures were 39 against 30. Cohere calls this up to 1.4x higher throughput. Long documents are a stronger point: the model scores 48.9 when translating two book chapters in one call, while Google Translate scores 21.3 and Gemma 4 31B scores 19.4. Quality is measured per paragraph with xCOMET-XL. In Cohere's cost chart, the model scores 80.1 at $0.000676 per task, averaging 661 tokens. Gemini 3.1 Pro Preview (high) costs $0.038928 per task, about 58 times more. Qwen 3.5 397B A17B costs $0.004525 per task and Command A+ costs $0.005158.

The fastest way to run the model is Cohere's Chat V2 API, where it is free until rate limits, using the model identifier north-small-translate-1-0. For self-hosting, Cohere publishes three checkpoints, the same ones it serves in production: a BF16 checkpoint runs on four B200 or eight H100 GPUs, an FP8 checkpoint runs on two B200 or four H100, and an NVFP4 W4A16 checkpoint runs on one B200 or two H100.