What is North Mini Code?
North Mini Code is a mixture-of-experts model that Cohere trained for agentic coding: repo-level changes in harnesses such as SWE-Agent and OpenCode, shell tasks, code reviews and plain code generation. It reads and writes text only. Cohere trained it across several agent harnesses rather than one, and it is Cohere's current and only coding model.
Its draw is size. Only 3B parameters work on each token, so it runs quickly on a single H100 or, in Cohere's 4-bit build, on about 20 GB of memory. The trade-off is quality: Cohere's own table shows Qwen3.6 35B-A3B, the obvious rival of the same size, ahead on most coding tests.
How does North Mini Code compare with Qwen3.6 35B-A3B?
It scores lower on five of the six coding tests in Cohere's own launch table and leads only on SciCode.
- Terminal-Bench v2
- 36.0 for North Mini Code against 51.5 for Qwen3.6 35B-A3B.
- SWE-bench Verified
- 67.6 against 73.4, and 40.2 against 49.5 on SWE-bench Pro.
- LiveCodeBench v6
- 70.3 against 80.4 on single-turn code generation.
- SciCode
- 38.2 against 35.8, the one test where North Mini Code is ahead.
Who is North Mini Code best for?
North Mini Code is best for developers and teams who want a free, Apache 2.0 coding agent model to run on their own GPU or Mac inside tools such as OpenCode.
Coding on your own machine
Only 3B of its 30B parameters work on each token, and Cohere's 4-bit build needs about 18 to 20 GB, so it runs on local GPUs and Macs.
Agent harness users
It was trained against several harnesses, so it works in tools such as OpenCode and SWE-Agent instead of being tuned to a single one.
Teams that keep code in-house
The Apache 2.0 weights can run on-prem or in a private cloud, so your source code never has to reach an outside API.
Repo changes and code reviews
It handles repo-level edits, shell-driven tasks and code reviews, with a 256K-token context and up to 64K tokens of output.
What's included with North Mini Code?
Training across several agent harnesses, thinking between tool calls, BF16, FP8 and 4-bit builds, faster serving, the Cohere API and managed hosting.
- Trained across agent harnesses
- Cohere mixed SWE-Agent, mini-SWE-agent, OpenCode and Terminus 2 styles into training, so it copes with the tool formats of different agents.
- Thinking between tool calls
- It reasons between steps, and Cohere asks you to pass that thinking back on each turn for the best results.
- Three weight builds
- BF16, FP8 and a 4-bit build that Cohere says keeps more than 99% of the full model's accuracy.
- Faster serving
- Cohere measured up to 2.8x the output throughput of Devstral Small 2 on the same hardware, and ships an EAGLE draft model for speculative decoding in vLLM.
- Cohere API
- Callable as north-mini-code-1-0 on Cohere's Chat V2 API, with tool use and structured outputs.
- Model Vault hosting
- Cohere can run it for you on a dedicated, managed instance billed by the hour.
How much does North Mini Code cost?
North Mini Code is free: the weights cost nothing to download, and Cohere's API serves it at $0 until you hit rate limits. Dedicated hosting on Model Vault costs from $7.50 an hour.
- Runs through the Cohere API as north-mini-code-1-0
- Trial keys allow 20 requests a minute and 1,000 calls a month
- 256K-token context, up to 64K tokens of output
- Open weights under the Apache 2.0 license
- BF16, FP8 and 4-bit versions on Hugging Face
- 30B parameters, 3B active per token
Cohere's docs say North Mini Code is free on both trial and production API keys until you reach the rate limits (20 requests a minute on a trial key, 500 on a production key). Trial keys may not be used for production or commercial work, and Cohere points production use to Model Vault, its dedicated managed hosting, where North Mini Code costs $7.50 an hour per instance on the L tier or $10.50 on the XL tier, usually after a waitlist. Monthly and annual Model Vault rates come from Cohere sales. Full pricing on cohere.com. We checked these prices on .
How do readers rate North Mini Code?
No one has rated North Mini Code yet. Be the first.
Sign in to rate. One rating per account.
North Mini Code pros and cons
It is fast, free and steady on tool calls for its size, but it trails Qwen3.6 35B-A3B on most of Cohere's own tests and setup can be fiddly.
Pros
- Quick on modest hardware: users report about 50 to 60 tokens a second on a MacBook with Cohere's 4-bit build, and Cohere measured up to 2.8x the output throughput of Devstral Small 2.
- Steady in agent tools: users report no failed tool calls, including in OpenCode, and no looping on Cohere's own 4-bit build.
- Hacker News and Reddit users call it surprisingly capable for its size and competitive with many Chinese open models.
- Apache 2.0 weights, free for commercial use and on-prem hosting; users note it suits organisations that rule out models trained in China.
Cons
- Cohere's own table puts it behind Qwen3.6 35B-A3B on five of six coding tests, including 36.0 against 51.5 on Terminal-Bench v2, and many Reddit and Hacker News users rank it below Qwen3.6 too.
- Some users report thinking loops and hallucinations, especially with community quantized builds.
- Setup can be fiddly: Cohere's model card asks for vLLM's main branch plus its melody library for correct parsing, and some early GGUF builds did not work.
- Text only, and the free API is rate limited and not for production; production means Model Vault from $7.50 an hour or your own GPUs.
Is North Mini Code open source?
Yes. North Mini Code's code is public under the Apache 2.0 licence, so you can run it on your own server.
What are the best North Mini Code alternatives?
North Mini Code is an open source alternative to KAT-Coder-Pro V2 and Composer 2.5.
KAT-Coder-Pro V2PaidKwaiKAT's agentic coding model for large enterprise projects and SaaS integration, served through StreamLake.
Composer 2.5FreeCursor's fast agentic coding model with a 200K context, in Standard and Fast tiers.
Qwen3-Coder-NextOpen SourceChoose it when you want a pay-as-you-go API for production as well as open weights; Alibaba charges from $0.30 / $1.50 per million tokens, and it also has 3B active parameters, in an 80B model.
Nemotron 3.5 LightningOpen SourceChoose it when you need more context: NVIDIA's model is the same 30B size with 3B active and supports up to 1M tokens, though under its OpenMDW licence rather than Apache 2.0.
Qwen3.6-27BOpen SourceChoose it when you need image or video input; it is a dense 27B model under Apache 2.0 with a 262K context that stretches to about 1M tokens, but slower per token than a 3B-active model.
Questions about North Mini Code
What is North Mini Code?
North Mini Code is Cohere's first coding model: a free, Apache 2.0 open-weight model with 30B parameters, 3B of them active, built to drive coding agents on your own hardware.
Who is North Mini Code for?
North Mini Code is best for developers and teams who want a free, Apache 2.0 coding agent model to run on their own GPU or Mac inside tools such as OpenCode.
Is North Mini Code free?
Yes. North Mini Code has a free API key plan: runs through the Cohere API as north-mini-code-1-0.
Is North Mini Code open source?
Yes. North Mini Code's code is public under the Apache 2.0 licence, so you can run it on your own server.
What are the best North Mini Code alternatives?
The closest alternatives on Recaply are KAT-Coder-Pro V2, Composer 2.5, Qwen3-Coder-Next, Nemotron 3.5 Lightning and Qwen3.6-27B.
Where can I try North Mini Code?
On its official website, cohere.com.
Can I use North Mini Code through the Cohere API in production?
Not on a trial key. Cohere's docs say North Mini Code is free on trial and production keys until you hit rate limits, but trial keys are barred from production and commercial use, and Cohere points production work to Model Vault, its dedicated hosting from $7.50 an hour.
What hardware do I need to run North Mini Code?
Cohere lists one H100 GPU at FP8 as the minimum. The full BF16 weights take about 61 GB, while Cohere's 4-bit build needs about 18 to 20 GB and runs on Hopper and Ada GPUs, and on Macs through MLX.
What license does North Mini Code use?
North Mini Code is released under the Apache 2.0 license, which allows commercial use. Cohere publishes BF16, FP8 and 4-bit weights on Hugging Face, so you can run it locally or on your own servers without paying Cohere.
Which coding tools work with North Mini Code?
Cohere trained it for OpenCode and tested it in SWE-agent, mini-SWE-agent and Terminus 2, and says it works with most coding agents. For local use, Cohere publishes an OpenCode config for a vLLM server that turns on interleaved reasoning.
Our verdict
North Mini Code is a good free pick if you want an Apache 2.0 coding agent model you can host yourself, but it is not the strongest model of its size.
Its 3B active parameters make it quick on one GPU or a Mac and users find it steady on tool calls, yet Cohere's own table puts Qwen3.6 35B-A3B ahead on five of six coding tests.
- Skip it for
- You only want the top benchmark score at this size, or you need image input.
- Check first
- Run Cohere's own FP8 or 4-bit build on a current runtime, since users report loops with some community builds.
Kimi K2.7 Code
Cohere Embed 5
MiniMax M2.1
SWE-grep
Qwen3.8-27B
Chroma Context-1