01
Overview
Cohere released North Small Translate 1.0 on September 10, 2026 as its first translation-specialized model in the North family. It is an open-weight sparse Mixture-of-Experts model built specifically for machine translation across 50 languages, with multiple deployment paths for research, API use and licensed enterprise hosting.
02
What is North Small Translate?
North Small Translate is a decoder-only sparse MoE model with 218 billion total parameters and 25 billion active parameters per token. Cohere documents 128 experts, eight selected per token, plus shared experts. It supports 16K input tokens and up to 16K output tokens.
The model is text-in/text-out and is optimized for translation rather than general chat. Cohere says it covers 50 languages including English, Arabic, Hindi, Bengali, Punjabi, Tamil, Telugu, Japanese, Korean, Vietnamese, Indonesian, Thai, Traditional Chinese and Simplified Chinese, plus major European languages.
03
Release and availability
The model launched publicly on September 10, 2026. Cohere provides several ways to use it:
- Download official weights from Hugging Face for research and non-commercial use.
- Use Cohere's hosted Chat API with model ID north-small-translate-1-0.
- Deploy it commercially through Cohere Model Vault.
- Access an enterprise localization path through RWS Language Weaver.
Cohere's current model documentation says API usage is free until rate limits are reached. That should not be interpreted as a permanent unlimited free tier; rate limits and commercial terms can change.
04
Open weights and license
The official weights are released under CC BY-NC 4.0 with Cohere Labs' Acceptable Use Policy. That makes the public checkpoint suitable for research and non-commercial use, but not a blanket commercial-use license.
Teams that need commercial deployment should use Cohere's commercial licensing options rather than assuming the Hugging Face license covers production use.
05
Model architecture
North Small Translate uses a 218B-parameter sparse MoE design with 25B parameters active per token. The attention stack mixes sliding-window attention with global attention in a 3:1 ratio. Cohere's model card lists a 4,096-token sliding window and 16K total context.
The important practical point is that active-parameter compute is much smaller than the total model size, but self-hosting still requires enough memory to hold the full checkpoint or an appropriate quantized form.
06
Hardware and quantization
Cohere publishes multiple official weight formats, including BF16, FP8 and NVFP4-style quantized variants. Its launch post lists minimum configurations as low as one B200 or two H100 GPUs for its W4A4 configuration. Cohere's current model documentation separately lists suggested hardware of two H100s or one B200.
Those numbers are official deployment references, not universal guarantees for every runtime, batch size, concurrency level or context length. Memory and throughput depend heavily on precision, framework and workload.
07
API model ID
The hosted model identifier is north-small-translate-1-0. Cohere documents support through its Chat API and Playground. Developers should confirm current endpoint behavior and rate limits before production deployment.
08
Benchmark results
Cohere reports a WMT26 All Languages score of 83.60 for the standard model and 84.36 for an agentic multi-pass translation workflow that attempts to identify and correct translation errors. Cohere's launch material compares those results with systems including DeepL NextGen, Google Translate, Gemma 4 31B and other open models.
These numbers are Cohere-run evaluations. They are useful evidence about how Cohere positions the model, but they should not be presented as independent proof that North Small Translate is universally better for every language pair or domain.
Cohere also reports stronger long-document translation performance and up to 1.4× higher output throughput than Gemma 4 31B in its internal tests. Again, these are vendor-reported figures and should be labeled accordingly.
09
Agentic translation
Cohere also describes an agentic translation mode that performs an additional pass to find and correct translation errors. In Cohere's own WMT26 evaluation, this increases the score from 83.60 to 84.36.
This does not mean the base model independently runs an autonomous agent. It is better understood as a multi-pass workflow around the translation model.
10
Pricing
Hosted API documentation currently says North Small Translate is free until rate limits are reached. For commercial dedicated deployment, Cohere's Model Vault pricing is instance-based rather than a simple token price. Its current standard pricing table lists North Small Translate at $57.50 for the relevant flex instance tier, with broader commitment pricing available through Cohere.
Because the hosted API, Model Vault and RWS paths use different commercial models, there is no single universal 'North Small Translate price.'
11
Who is it for?
The model is most relevant to teams that need multilingual translation at scale, want sovereign or self-controlled deployment, need open weights for research, or want to integrate translation into developer and enterprise workflows without relying entirely on a closed translation API.
It is especially interesting for localization, multilingual content operations, customer-support translation, international documentation and high-volume business translation workflows.
12
Important limitations
The public weights are non-commercial by default. Cohere's strongest benchmark claims are vendor-run. The model is large even though only 25B parameters are active per token. Exact hardware requirements vary by checkpoint and runtime. Hosted API rate limits can change. Translation quality will differ by language pair, domain, terminology and input length.
13
FAQ
Is North Small Translate open source?
The more precise description is open-weight. Cohere publishes model weights under CC BY-NC 4.0 plus its acceptable-use terms, which restrict commercial use.
How many languages does it support?
Cohere documents 50 languages.
What is the model size?
218B total parameters with 25B active per token.
What is the context window?
16K input and 16K maximum output tokens.
Is there an API?
Yes. Cohere documents model ID north-small-translate-1-0 in its Chat API.
Is the API free?
Cohere currently says the model is free until rate limits are reached. Recheck the official documentation before budgeting production use.
Can I run it locally?
Yes, official weights and quantized checkpoints are available, but this is still a very large model. Cohere documents configurations using B200 or H100 GPUs rather than consumer-grade hardware as the baseline deployment target.
Sources
Primary and supporting sources
Facts were rechecked against the linked sources immediately before publication. Pricing, product availability and rollout status can change.