The notable thing about Cohere’s North Mini Code isn’t that it writes code — it’s the shape they picked: 30B total parameters, 3B active, mixture-of-experts, Apache 2.0. That’s a deliberate bet that the agentic coding market splits along deployment lines, not leaderboard lines.

Most coding models optimize for the benchmark and assume you’ll rent inference from someone else’s GPUs. A 3B-active MoE optimizes for the opposite — running where the code already lives: inside a VPC, on a single accelerator, behind an air gap. Cohere frames this as “sovereign AI,” which in enterprise terms means the model your security and compliance teams will actually sign off on. In production AI work the blocker is rarely raw capability; it’s where the weights sit and who gets to see the prompts and the source they touch.

The 3B-active figure is the operational story. An agentic coding task is not one call — it’s read-file, plan, edit, run-tests, re-plan, often dozens of forward passes before a diff lands. At that volume, active-parameter count is your latency and cost budget, not a footnote. A dense 30B that’s smart-but-slow can lose inside a loop to a sparse 30B that’s good-enough-but-fast, because iteration amortizes a lot of mediocrity that a single-turn eval would punish.

What I’d want before trusting it: eval numbers on the full loop, not the single turn — task completion across multi-step edits, test-pass rate after self-correction, and how badly quality degrades when you quantize to fit one GPU. The HN thread is already arguing about whether “open weights you can self-host” beats a frontier API you can’t audit. For regulated buyers, is a smaller model you fully control now worth more than a larger one you have to trust?