Cohere · 2026-06-09 · major
Cohere Ships North Mini Code 1.0 — Apache-2.0 Sparse 30B/3B-Active Mixture-of-Experts Coding Model With a 256K Context, 64K Max Output, 83.2% Pass@1 on SWE-Bench Verified, 63% on Terminal-Bench v2, and 2.8× the Output Throughput of Devstral Small 2
Cohere's first developer-focused open-weights model. 30B-parameter MoE with 3B active per token, 128 experts, 256K context, Apache 2.0, posted to Hugging Face with BF16 and FP8 checkpoints.
Cohere's first model aimed squarely at developers — a 30B/3B-active MoE coding model under Apache 2.0.
Key specs
| Parameters | 30B |
|---|---|
| Active params | 3B |
| Context window | 256K |
| Experts | 128 |
| Active experts | 8 |
| Max output | 64K |
| Swe bench verified pass@1 | 83.2% |
| Terminal bench v2 pass@1 | 63.0% |
| Mini swe agent pass@1 | 61.0% |
| Artificial analysis coding index | 33.4 |
| Throughput vs devstral small 2 | 2.8× |
What is it?
North Mini Code 1.0 is an open-weights coding model from Cohere Labs, trained for agentic software-engineering work like running tests, editing files, and driving terminals. The weights, a tokenizer, and an FP8 quantization land on Hugging Face under Apache 2.0, with hosted inference also offered via Cohere's API and OpenRouter.
How does it work?
It is a decoder-only Transformer with a sparse Mixture-of-Experts: 30B total parameters, 3B active per token, 128 experts with 8 routed per token via a sigmoid router, SwiGLU activations, and interleaved sliding-window plus global attention in a 3:1 ratio. Post-training stacks cascaded supervised fine-tuning with asynchronous reinforcement learning with verifiable rewards (RLVR), using windowed FIFO queues and CISPO loss over 70k+ verifiable coding tasks. Context length is 256K tokens with up to 64K tokens of generation.
Why does it matter?
It is the smallest open-weights model so far to clear 80% on SWE-Bench Verified and 60% on Terminal-Bench v2, beating Devstral Small 2, Gemma 4 26B, and Qwen3.5 35B-A3B on the Artificial Analysis Coding Index while serving 2.8× the output throughput of Devstral Small 2. For teams that want an agentic coding model they can self-host without buying Anthropic or OpenAI capacity, this lowers the bar substantially.
Who is it for?
developers running self-hosted coding agents and teams looking for an Apache-2.0 alternative to Devstral, Gemma 4, and Qwen3.5
Try it
huggingface.co/CohereLabs/North-Mini-Code-1.0