Cohere has released North Mini Code, a 30B-parameter Mixture-of-Experts model with 3B active parameters built specifically for agentic software engineering — the first in the company’s new family of developer-facing models. The weights are live on Hugging Face under the Apache 2.0 license, a direct bet that open, efficient coding models can compete with far larger frontier systems.
The Release
- Architecture: sparse MoE transformer with 128 experts, 8 active per token, interleaved sliding-window and global self-attention in a 3:1 ratio
- License: fully open-weight under Apache 2.0
- Weights:
North-Mini-Code-1.0(BF16) andNorth-Mini-Code-1.0-fp8(quantized) - Availability: Hugging Face, the Cohere API, and the OpenCode coding agent
North Mini Code is Cohere’s answer to the same question the rest of the industry is chasing: how to build a model that is fast, cheap, and reliable enough to serve as the backbone of real coding agents.
Post-Training for Coding Excellence
The team used a two-stage cascaded supervised fine-tuning (SFT) pipeline followed by reinforcement learning with verifiable rewards (RLVR). The first SFT stage mixes 70% code tokens with reasoning and instruction-following data; the second stage primes the model on a 4.5B-token diet of agentic and reasoning samples covering over 70,000 verifiable tasks across roughly 5,000 unique repositories.
Training environments are containerized and deduplicated against SWE-Bench and SWE-Bench-Pro sources to prevent eval leakage — a level of hygiene that makes the reported numbers unusually trustworthy.
Benchmark Results
After SFT, North Mini Code posts 80.2% pass@10 on SWE-Bench Verified and 55.1% pass@10 on Terminal-Bench v2. On Artificial Analysis’ Coding Index it scores 33.4, ahead of Qwen3.5 (35B-A3B), Gemma 4 (26B-A4B), and Devstral Small 2 — and even beats substantially larger closed competitors like Mistral Small 4 and Devstral 2.
The subsequent RLVR run added 7.9% absolute pass@1 on Terminal-Bench v2 and 3.0% on SWE-Bench, while cutting invalid tool calls and repetitive looping.
Robustness Across Harnesses
Rather than tuning for a single scaffold, Cohere trained across multiple agent harnesses — SWE-Agent, mini-SWE-agent, and OpenCode — which cost only 6% of the SFT mix but delivered a 10% gain on the OpenCode harness. A hidden payoff: 61.0% pass@1 on mini-SWE-Agent emerged for free through cross-harness transfer.
What This Means
North Mini Code lands right as enterprise teams demand open models that can actually drive terminal agents without supervision. The 3B-active design keeps inference cheap, while Apache 2.0 removes licensing friction. If the market has learned anything this year, it’s that the coding-agent crown goes to whoever ships the cheapest reliable model — and Cohere just put itself firmly in that race.