Cohere North Mini Code
Cohere's first open-source agentic coding model: a 30B Mixture-of-Experts architecture that activates only 3B parameters per forward pass, runs on a single H100 GPU, and is licensed Apache 2.0. The model is purpose-built for coding agent workflows — code generation, review, architecture design, and sub-agent coordination — achieving 33.4 on the Artificial Analysis Coding Index with 2.8× the throughput of Devstral Small 2. It's the "sovereign developer" play: a model you can run on your own hardware with no API dependency.
Key Quotes
"Small, efficient, and open-source — our first agentic coding model, built for the sovereign developer ecosystem." — Cohere team blog post
"Built to run where you need it" — positioning North Mini Code as a local-first alternative to cloud-only coding models, not just a cheaper API option.
The "sovereign" framing is doing real work here. This isn't just about cost — it's about developers who can't or won't send proprietary code to a third-party API. Cohere is betting that market is larger than the cloud-only one.
Key Themes
Architecture
MoE with extreme sparsity: 30B total, 3B active. That's a 10:1 sparsity ratio — more aggressive than Mixtral (8:1) and closer to DeepSeek-V2's design philosophy. The bet is that coding tasks benefit from specialized experts more than general reasoning tasks do.
256K context, 64K generation: Large enough for full-repo context in most codebases, though the 64K generation ceiling means it won't produce novel-length diffs. Practical for real-world PRs.
Single H100 at FP8: This is the headline. A 30B MoE fitting on one consumer-adjacent GPU means this is genuinely runnable by individual developers, not just enterprises. At 4-bit quantization it reportedly fits on a Mac Studio (~20GB).
Benchmark Reality Check
The 33.4 on the Coding Index puts it above GLM-4.7-Flash (25.9) but below Qwen3.6 35B (35.2). That's the right neighborhood for a 3B-active model — punching above its weight class, but not magically outperforming larger dense models.
The throughput claims (2.8× Devstral Small 2) matter more in practice. For agentic loops where the model makes dozens of sequential calls, per-token latency compounds into wall-clock time that determines whether the agent feels responsive or glacial.
The non-coding weakness (14% GDPval-AA, 37% τ²-Bench Telecom) is honest and expected. This is a specialist, not a generalist. Compare to MiMo Code which takes the opposite approach — general capability with specialized harness design.
Critical Analysis
Strong: The Apache 2.0 license is the real story. Cohere's previous models used CC-BY-NC, which meant "open weights but don't use them commercially." Apache 2.0 removes that friction entirely. For the ecosystem of tools like Pi, clawdBot, and Odysseus that let developers run agents locally, this is a genuinely useful new option. The single-H100 requirement puts it within reach of individual developers with a decent GPU budget.
Strong: The MoE sparsity ratio (30B→3B) is architecturally interesting. It suggests Cohere believes coding tasks have high expert specialization — different kinds of code (Python vs Rust, backend vs frontend, generation vs review) activate different pathways. If true, this is a more efficient use of parameters than a dense 3B model. If the specialization doesn't hold up in practice, it's just a 3B model with 30B of overhead.
Weak: "Chatty" output is a real problem for an agentic coding model. Verbose responses waste context and slow down agent loops — every extra token the model generates is a token the next turn has to process. The 2.8× throughput advantage could be partially or fully eaten by verbosity. Compare Honey I Shrunk the Coding Agent's finding that harness design matters more than raw model performance — a chatty model with a bad harness is worse than a terse model with a good one.
Weak: The 14% on GDPval-AA and 37% on τ²-Bench Telecom suggest the model falls apart outside its training distribution. For a coding agent that will inevitably encounter non-coding tasks (reading docs, understanding user intent, debugging logic errors that require domain knowledge), this brittleness is a liability. The Components of a Coding Agent taxonomy makes clear that coding agents need more than code generation — they need reasoning, planning, and tool use that may draw on capabilities North Mini Code simply doesn't have.
The bigger picture: This is part of a wave. Between North Mini Code, Qwen3.6, MiMo Code, and the 3B-to-9B models documented in Honey I Shrunk the Coding Agent, we're seeing a convergence: coding agents are becoming small enough to run locally, and the open-weight models are catching up fast. The question isn't whether local coding agents will be viable — it's whether the cloud models' advantage (bigger context, faster iteration, no hardware management) justifies their dependency cost. For the "sovereign developer" Cohere is targeting, the answer is already no.
Sources: Cohere Open-Sources North Mini Code: A Coding Agent That Runs on a Single H100, Cohere blog: Introducing North Mini Code Original VentureBeat article: https://venturebeat.com/technology/cohere-open-sources-a-coding-agent-that-runs-on-a-single-h100 (paywalled) Last updated: 2026-06-15