Cohere North Small Translate is a 25B-active open-weight MoE built for translation across 50+ languages
The dedicated translation model supports 16K input and output, with research and non-commercial access under CC BY-NC 4.0.
Key takeaways
Cohere released North Small Translate on September 10, an open-weight mixture-of-experts model dedicated to machine translation across more than 50 languages. The model has 218 billion total parameters and 25 billion active parameters, with 16K input and 16K output context. Cohere lists one B200 or two H100 GPUs at W4A4 as minimum configurations. In vendor-run evaluations using GPT-5.6-Sol as a judge, the standard model scored 83.60 across WMT26 languages and an agentic variant scored 84.36. Cohere also reports up to 1.4 times the output throughput of Gemma 4 31B under identical hardware and concurrency. The model card says the non-commercial CC BY-NC 4.0 license carries an acceptable-use addendum and requires Cohere Labs AUP compliance; commercial production requires a separate commercial license.
Direct answer
Cohere released North Small Translate on September 10, an open-weight mixture-of-experts model dedicated to machine translation across more than 50 languages. The model has 218 billion total parameters and 25 billion active parameters, with 16K input and 16K output context. Cohere lists one B200 or two H100 GPUs at W4A4 as minimum configurations. In vendor-run evaluations using GPT-5.6-Sol as a judge, the standard model scored 83.60 across WMT26 languages and an agentic variant scored 84.36. Cohere also reports up to 1.4 times the output throughput of Gemma 4 31B under identical hardware and concurrency. The model card says the non-commercial CC BY-NC 4.0 license carries an acceptable-use addendum and requires Cohere Labs AUP compliance; commercial production requires a separate commercial license.
Verified facts
Cohere lists a 218B-total, 25B-active MoE architecture with 16K input, 16K output, and support for more than 50 languages.
The model card applies a non-commercial CC BY-NC 4.0 license, an acceptable-use addendum, and the Cohere Labs AUP; commercial production requires a separate commercial license.
Cohere's WMT26 evaluation uses GPT-5.6-Sol as judge and reports 83.60 for the standard model and 84.36 for the agentic variant, plus up to 1.4x throughput in internal testing.
What changed
- A dedicated translation model covers more than 50 languages and long input-output windows
- Twenty-five billion active parameters separate total model size from per-step compute
- Standard and self-correcting agentic variants receive separate reported scores
- Downloadable open weights are explicitly separated from commercial usage rights
Impact for AI users
For cross-border products, a dedicated translation model allows terminology, long-document consistency, throughput, and private deployment to be evaluated together. A single all-language average is insufficient because language region, document length, domain vocabulary, and human quality criteria can change the outcome. Licensing is also a release prerequisite: the non-commercial license, acceptable-use addendum, and AUP all require review, while commercial production needs a separate commercial license and a full cost model covering hardware, concurrency, and qualified review.
Operating checklist
- Build a language-by-domain matrix for target markets, including glossaries, long documents, markup, and difficult samples that require human judgment.
- Blind-test the standard, agentic, and incumbent translation systems, recording quality, latency, throughput, and revision counts separately.
- Validate memory, concurrency, long-document degradation, and quantization error on the target W4A4 hardware instead of adopting vendor throughput directly.
- Obtain a separate commercial license before production use, then verify the acceptable-use addendum, Cohere Labs AUP, weight provenance, and qualified-review requirements.
AI frontier news and analysis, AI software and model tools, AI skill tutorials and validation methods, and AI account and permission guidance
FAQ
Do open weights mean unrestricted free commercial use?
No. The model card requires the non-commercial CC BY-NC 4.0 license, an acceptable-use addendum, and Cohere Labs AUP compliance; commercial production needs a separate commercial license.
Are the 83.60 and 84.36 scores independent third-party results?
No. They are Cohere-published WMT26 evaluations that use GPT-5.6-Sol as the judge.
Can the model run on one ordinary consumer GPU?
Cohere lists one B200 or two H100 GPUs at W4A4 as minimum hardware and makes no claim that a consumer GPU meets that requirement.
Summary
North Small Translate combines multilingual quality, long-document throughput, and private deployment in one specialized model, but adoption still depends on per-language evidence, target hardware, and commercial licensing.
This AI-assisted article is checked by ENHE AI automation for official sources, bilingual fields, media rights, page safety, and historical duplication before publication.
What this means for everyday users
For cross-border products, a dedicated translation model allows terminology, long-document consistency, throughput, and private deployment to be evaluated together. A single all-language average is insufficient because language region, document length, domain vocabulary, and human quality criteria can change the outcome. Licensing is also a release prerequisite: the non-commercial license, acceptable-use addendum, and AUP all require review, while commercial production needs a separate commercial license and a full cost model covering hardware, concurrency, and qualified review.
Tools you may use
Related tutorials
[[zh]][[/zh]][[en]]Review a public website[[/en]]
SearchScope SEO/GEO
[[zh]][[/zh]][[en]]Access and usage guide[[/en]]
AI Prompt Management System | 418 Bilingual Prompts for Writing, SEO and AI Creation
High-Frequency AI Prompts for Work, Learning, and Teaching guide
High-Frequency AI Prompts for Work, Learning, and Teaching
Related Tools And Tutorials
Use the following ENHE AI sections to continue from the news signal into tool selection, account-service guidance, or practical learning.
Related reading
Microsoft and NVIDIA Publish an Open-Weight Model Risk Guide With Five Risk Categories
Microsoft and NVIDIA published their first joint resource for enterprise open-weight model risk management on October 16, 2026. The guidance organizes discussion into five broad categories: cybersecurity, data protection, content, human-AI interaction, and bias. Its central point is that weight openness alone does not determine risk; the model, intended use, deployment architecture, users, data, tools, and safeguards all shape the actual system. It also distinguishes model access from deployment choice and warns that self-operated deployments can transfer controls from a managed provider to the enterprise. This article turns the framework into a practical selection, ownership, evaluation, and review checklist for ordinary teams.
Reflection Beam Arrives on AWS: How to Choose an Agentic Coding Deployment
AWS introduced Reflection AI's Beam on October 16, 2026 as an open-weight model for agentic coding and long-horizon work. Beam is designed to resolve software issues, modify code, call tools, interpret intermediate results, and continue multi-step workflows. AWS describes a sparse mixture-of-experts architecture with 501 billion total parameters, 23 billion active parameters, and an effective one-million-token context length reported by Reflection. Weights are released under Apache 2.0 in BF16, FP8, and NVFP4 variants. This guide separates the managed SageMaker JumpStart path from the more controllable HyperPod route and explains the operational evidence teams should collect before trusting either in real projects.
Kimi K3 Reaches GitHub Copilot: Check Plan Access, Model Policy, and Usage Billing
GitHub announced on August 6, 2026 that Kimi K3 is gradually rolling out to Copilot Pro, Pro+, Max, Business, and Enterprise plans. Eligible users can select it in places such as Visual Studio Code, Copilot CLI, GitHub, and supported IDEs, but availability depends on plan, client, and rollout status. GitHub says Kimi K3 uses provider list pricing under usage-based billing. Business and Enterprise administrators must enable the Kimi K3 policy before members can use it and should review open-weight model governance. For ordinary users, the practical first step is a bounded non-production test with a recorded budget, permissions, changes, tests, and human review rather than an immediate production rollout.
NVIDIA Launches Open Secure AI Alliance as Open AI Agents Move Toward Auditable Collaboration
NVIDIA and a group of AI and infrastructure organizations launched the Open Secure AI Alliance on July 27, 2026 and highlighted the open-source NOOA agent framework. For ordinary users and teams, the practical lesson is not to treat open source as an automatic security guarantee. A deployable agent should expose its model choice, Python agent code, tool permissions, dependencies, traces, approval steps, and containment boundary. NOOA supports familiar testing, tracing, refactoring, and version-control workflows, but its repository also warns that in-process validation is not a security boundary when agents execute model-generated code. Use non-sensitive data and operating-system-level isolation before granting real accounts, files, publishing rights, or payment access.
GitHub Brings Kimi K2.7 Code Into Copilot's Main Model Picker
GitHub announced on July 1, 2026 that Kimi K2.7 Code is generally available in GitHub Copilot and is the first open-weight model selectable in the Copilot model picker. For ordinary AI users, the practical signal is not simply that another coding model has arrived. It means model choice, AI credits, provider pricing, Azure hosting, and organization-level policy are becoming part of the same daily coding workflow. Teams using Copilot should treat Kimi K2.7 Code as a lower-cost coding option to test, not as an automatic replacement for every model. The sensible next step is to compare tasks, permission boundaries, output quality, and cost behavior before enabling it broadly.
How to Test Kimi K2.7 Code in Copilot Safely
Testing Kimi K2.7 Code inside Copilot should be treated as a controlled workflow, not a casual switch. Start by confirming whether the model is available in your plan and whether administrators have enabled it. Then use a sample repository, read-only tasks, and low-risk prompts such as explaining code, writing tests, or suggesting small fixes. Track AI-credit usage and compare the output with your normal Copilot model. Do not send secrets, proprietary customer data, or production credentials. The goal is to decide whether the model is useful for a defined coding workflow, not to prove that one model should replace all others.