Cohere launches North Small Translate: MoE translation model with 25B active and 218B total parameters

Cohere and its research arm Cohere Labs this week released North-Small-Translate-1.0, a dedicated translation model built on a sparse Mixture-of-Experts architecture with 25 billion active parameters drawn from 218 billion total. The model targets high-quality translation across 50 languages and is distributed as open weights under a CC BY-NC 4.0 license that restricts commercial use and requires compliance with Cohere Labs' acceptable-use policy. Access to the files on Hugging Face is gated behind agreement to share contact details and receive email updates from the company.

Architecture and production deployment

The sparse MoE choice lets Cohere keep the inference cost of a 25B model while drawing on the knowledge capacity of a far larger one. Context length sits at 16 thousand tokens for both input and output — enough for long business documents but unremarkable compared with general-purpose models of the same generation. The three quantizations available on Hugging Face are the exact checkpoints Cohere runs in production, and the company recommends greedy decoding only, the same setting it uses internally. Minimum memory requirements start at 8×H100 GPUs with 70GiB per card, and loading requires an explicit `max_memory` setting to prevent a crash during MoE layer fusion.

Output format and dedicated parsing library

The model wraps its answer in structural markers `<|START_TEXT|>` and `<|END_TEXT|>` that are not registered as special tokens, a deliberate decision so that vLLM's reasoning parsers and tool-call parsers can see them. The practical consequence: `skip_special_tokens=True` does not strip them, and the raw output includes the system instructions as well. Cohere supplies the `cohere_melody` library — the same parser vLLM uses — for extracting clean text, and demonstrates a dedicated chat template that feeds the translation instruction and extracts only the newly generated tokens.

License, access and an important distinction

The CC BY-NC 4.0 license makes the release "open weights" rather than open source, a distinction Cohere itself insists on. The commercial restriction and mandatory pre-registration place the model in a different category from Meta's NLLB or Nvidia's translation models, which are available under more permissive licenses. On the other hand, the fact that the quantizations are identical to those used in production gives developers a realistic performance reference point, with no gap between a "community" version and an "enterprise" version.