Cohere's new open-source coding agent runs on a single H100 — but it's three times more verbose than comparable models

North Mini Code is purpose-built for agentic software engineering and deployable on commodity hardware. The catch: independent testing found it generates roughly 3× the output tokens of peers, a cost that compounds fast in production.

Written by Lena Armitage · Bureau Tech · Jun 14, 2026

The surprising number isn't the parameter count

Cohere's North Mini Code, released Tuesday under an Apache 2.0 license, fits a capable coding agent onto a single NVIDIA H100 GPU. That's the headline. The number worth scrutinizing is different: according to independent benchmarking by Artificial Analysis, the model generated **75 million output tokens** to complete the Intelligence Index evaluation. The class median was 25 million.

For a model positioned as a cost-efficient alternative to managed APIs, that verbosity is a material caveat — one the benchmark leaderboard rankings don't surface on their own.

What the model actually is

North Mini Code is a sparse **mixture-of-experts (MoE)** model, an architecture where only a subset of the model's parameters activate for any given input. It has 128 experts total, with 8 active per token. Total parameter count is 30 billion; active compute at inference is closer to a 3-billion-parameter model. That's why it fits on a single H100 — and why Cohere co-founder Nick Frosst was able to demo it running on a Mac Studio using roughly 20 GB of RAM.

The model supports a 256,000-token context window with a maximum generation length of 64,000 tokens, and it's available on Hugging Face.

Built for agents, not adapted for them

The distinction Cohere is drawing — and it's a meaningful one — is between models fine-tuned for code generation and models trained from the ground up for **agentic workflows**: multi-step tasks involving tool calls, shell interactions, sub-agent orchestration, and code review across large codebases.

Cohere trained North Mini Code through two stages of supervised fine-tuning followed by reinforcement learning with verifiable rewards, using more than 70,000 verifiable tasks spanning approximately 5,000 repositories (deduplicated against SWE-Bench, a standard software engineering evaluation). Crucially, training ran across three distinct agent scaffolds — SWE-Agent, Mini-SWE-Agent, and OpenCode — rather than optimizing for a single harness. Cohere reports a 10-percentage-point gain on OpenCode evaluation from that multi-harness approach.

Where the benchmarks land — and where they don't

Artificial Analysis independently ranks North Mini Code **8th of 127 comparable open-weight models** on output speed at 210 tokens per second, with a time-to-first-token of 0.25 seconds against a class median of 1.95 seconds. It places 18th of 127 on the Artificial Analysis Intelligence Index.

Cohere's own reported benchmarks claim it outperforms open-source models up to four times its parameter count — including models at 120 billion parameters. That's a strong claim. It has not yet been independently replicated, and vendor-reported benchmark comparisons have a well-documented history of favorable framing. The speed numbers from Artificial Analysis are more trustworthy precisely because they're third-party.

The verbosity finding from the same Artificial Analysis data deserves equal weight: three times the output tokens of comparable models is not a rounding error. In high-volume agentic pipelines, it compounds directly into inference cost and latency.

The real decision for enterprise teams

North Mini Code enters a market that includes Mistral Devstral Small 2, GitHub Copilot, Cursor, and Anthropic's Claude — each with different cost and deployment profiles. Anthropic's Claude, currently the most capable publicly available managed coding model, runs at $50 per million output tokens with no on-premises option.

For teams running production agentic coding pipelines, the choice is now concrete rather than theoretical: self-hosted open-weight models on owned hardware versus managed APIs with their associated infrastructure overhead and data-residency constraints. North Mini Code makes the self-hosted side of that comparison more credible than it was a week ago.

The evaluation step that benchmark rankings skip is throughput testing against actual workload volume. Given the verbosity data, that step is not optional.

Key takeaways

North Mini Code is a sparse mixture-of-experts (MoE) model — 30 billion total parameters, 3 billion active per token — that fits on a single H100 and is freely licensed under Apache 2.0.
Cohere trained it specifically for agentic workflows using more than 70,000 verifiable tasks across ~5,000 repositories, rather than adapting a general-purpose base model.
Artificial Analysis independently ranked it 8th of 127 open-weight models on output speed (210 tokens/sec), but it generated 75 million output tokens on the Intelligence Index benchmark versus a class median of 25 million — a verbosity gap that inflates inference cost in high-volume pipelines.
Cohere's vendor-reported benchmarks claim it outperforms open-source models up to four times its parameter count; those claims have not been independently replicated and should be treated with caution until third-party evaluations catch up.
The practical choice for enterprise teams is now concrete: self-hosted North Mini Code on owned hardware versus managed frontier models like Claude at $50/million output tokens — a real architectural and cost decision, not a hypothetical one.

FAQ

What is a mixture-of-experts (MoE) model, and why does it matter for deployment?

A mixture-of-experts model routes each input token through only a subset of its total parameters — in North Mini Code's case, 8 of 128 experts activate per token. This means the active compute at inference is much smaller than the total parameter count suggests, which is why a 30-billion-parameter model can run on a single H100 GPU with roughly 20 GB of RAM.

What does 'Apache 2.0 license' mean for enterprise use?

Apache 2.0 is a permissive open-source license that allows commercial use, modification, and distribution without requiring derivative works to be open-sourced. For enterprise teams, it means North Mini Code can be deployed in production environments and integrated into proprietary pipelines without licensing fees or open-source disclosure obligations.

How does North Mini Code's verbosity affect real-world cost?

Artificial Analysis found the model generated 75 million output tokens on its Intelligence Index benchmark, versus a class median of 25 million. In self-hosted deployments, more output tokens mean longer inference time and higher GPU utilization per task. In any pipeline where tasks are high-volume or time-sensitive, that 3× verbosity gap translates directly into throughput constraints and operating cost.

How does it compare to Anthropic's Claude for coding tasks?

Claude is currently the highest-capability managed coding model publicly available, priced at $50 per million output tokens with no self-hosting option. North Mini Code ranks 18th on the Artificial Analysis Intelligence Index versus Claude's higher placement, so there is a capability gap — but for teams with data-residency requirements or high-volume workloads where per-token cost is the binding constraint, the tradeoff may favor the open-weight option. The verbosity difference means the effective cost comparison requires modeling actual output token volume, not just list price.

What is SWE-Bench, and why did Cohere deduplicate against it?

SWE-Bench is a widely used benchmark that tests models on real GitHub issues requiring code changes. Deduplicating training data against SWE-Bench means Cohere removed tasks that overlap with the benchmark's test set, which is standard practice to prevent inflated scores from training-set contamination. It's a methodological detail worth noting because not all model releases disclose it.

Citations

Cohere open-sources a coding agent that runs on a single H100North Mini Code is a 30-billion-parameter MoE model available under Apache 2.0, designed for agentic coding workflows and deployable on a single H100 GPU.
Artificial Analysis Intelligence Index — North Mini Code entryArtificial Analysis independently ranked North Mini Code 8th of 127 open-weight models on output speed at 210 tokens/sec and 18th on the Intelligence Index; it generated 75 million output tokens versus a class median of 25 million.
North Mini Code — Hugging Face model card and technical postCohere claims North Mini Code outperforms open-source models up to four times its parameter count on reported benchmarks, and reports a 10-percentage-point gain on OpenCode evaluation from multi-harness training.
Nick Frosst post on X regarding North Mini Code positioningCohere co-founder Nick Frosst described North Mini Code as 'small, cost effective, apache 2.0, and locally deployable' and contrasted it with 'large, expensive, proprietary and hegemonic' alternatives.