Cohere Labs' North Small Translate Beats DeepL and Google Translate Across 50 Languages

Cohere Labs just dropped North Small Translate, a Mixture-of-Experts model built specifically for machine translation across more than 50 languages. It nods to Cohere's origin story (several founders were on the original Transformer paper, which was itself motivated by Google Translate) and takes a fairly aggressive shot at the incumbents: on the WMT26 benchmark, it beats DeepL, Google Translate, and open-weight competitors like GLM 5.2 and Mistral Large 3.

A translation-tuned MoE, not a repurposed chatbot

North Small Translate is a decoder-only sparse Mixture-of-Experts Transformer with 25B active parameters and 218B total, using 128 experts of which 8 are activated per token, alongside shared experts applied to every token. The attention layers interleave sliding-window attention (window size 4096) with Rotary Positional Embeddings and global attention layers without positional embeddings, in a 3:1 ratio first introduced in Command A. The router uses a sigmoid activation over the expert logits and normalizes over the top-k selected experts.

Context is 16K input and 16K output, and the model was post-trained specifically for translation quality rather than being a general chatbot with a translation prompt bolted on. Cohere developed it in partnership with RWS, whose Language Weaver research team and language experts helped shape real-world performance during training.

The benchmark numbers

On the WMT26 benchmark averaged across all languages, North Small Translate scores 83.60, compared with 81.56 for Qwen 3.5 397B A17B, 76.50 for GLM 5.2 FP8, 81.37 for DeepL NextGen, 79.46 for Gemma 4 31B, and 68.20 for Google Translate. An agentic multi-pass variant that finds and fixes its own errors bumps that to 84.36.

Regional performance is where things get interesting. Both versions outperform DeepL NextGen across every non-European region tested, with the largest advantage in South Asia and MENA (roughly 8-10 points ahead of DeepL), a moderate edge in Southeast Asia (about 4-5 points), and the narrowest edge in East Asia (about 1-3 points). In Europe, it beats Gemma 4 31B by nearly 10 points on EU languages while running roughly even with it in South Asia.

Throughput and long documents

Speed is the other pitch. In Cohere's testing, North Small Translate hit up to 1.4x higher output throughput than Gemma 4 31B on identical hardware: 112 vs 81 output tokens per second at low concurrency and 39 vs 30 at high concurrency, or 30-38% more tokens per second.

Long context is often where translation models fall over. North Small Translate scores 48.9 on Cohere's long-context evaluation, more than double Google Translate (21.3) and Gemma 4 31B (19.4), measured by translating two chapters of a book in a single call and scoring each paragraph with xComet-XL.

What it costs to run

The weights are free to download but the hardware bill is real. Cohere ships three quantizations, all of which are the checkpoints they run in production:

• BF16 needs 4x B200 or 8x H100
• FP8 needs 2x B200 or 4x H100
• NVFP4 W4A16 (4-bit weights) fits on 1x B200 or 2x H100

For anyone with a commercial license through RWS's Language Weaver, Cohere pitches strong cost efficiency: an 80.1 score at $0.000676 per task using only 661 tokens on average, compared with Gemini 3.1 Pro Preview at $0.038928 per task.

Running it yourself

The model is on Hugging Face today under a CC BY-NC 4.0 license for research and non-commercial use. There is a hosted Space for kicking the tires, and it runs through standard transformers or vLLM. One gotcha: the model wraps its reply in structural markers like <|START_TEXT|> and <|END_TEXT|> that are deliberately not registered as special tokens, so skip_special_tokens=True does not remove them. Cohere recommends parsing output with their cohere_melody library, the same parser vLLM uses.

A minimal vLLM launch on 8x H100 looks like:

vllm serve CohereLabs/North-Small-Translate-1.0 \
  -tp 8 \
  --max-model-len 32768 \
  --tool-call-parser cohere_command4 \
  --reasoning-parser cohere_command4 \
  --enable-auto-tool-choice

Who this is actually for

If you are shipping a translation feature across many locales and you have been paying per-character DeepL or Google fees, or you have compliance reasons to keep the workload in-house, this is the most credible open option in that lane right now. The 50 supported languages cover most of Europe, the major Southeast and East Asian languages, and a reasonable slice of MENA and South Asia, so coverage is broad without pretending to be everything.

The catch is the non-commercial license. For research, evaluation, or internal experimentation the weights are yours; for a shipped product you are either talking to Cohere sales or accessing it through RWS's Language Weaver. That framing (open weights for the community, commercial channel for enterprise) is becoming the default posture for frontier labs releasing specialist models, and translation may be one of the cleaner examples: narrow enough scope that a purpose-built model can genuinely outperform GPT-class generalists, broad enough demand that enterprises will pay for a supported pipeline.