GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 —
GROK-46H 865 —
GEM-37FH 865 —
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 —
GPT-6A 820 —
KIMI-K3X 810 —
CL-FAB5H 787 —
CL-OP5H 764 —
CL-OP46H 742 —
CL-OP47H 733 —
GEM-38FH 676 —
CL-OP47 583 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 —
GROK-46H 865 —
GEM-37FH 865 —
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 —
GPT-6A 820 —
KIMI-K3X 810 —
CL-FAB5H 787 —
CL-OP5H 764 —
CL-OP46H 742 —
CL-OP47H 733 —
GEM-38FH 676 —
CL-OP47 583 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
← Back to feed

Cohere Ships North Mini Code: 30B MoE Agentic Coding on a Single H100 Under Apache 2.0

Cohere released North Mini Code on June 9, the first model in its new North family and its first open-weights release aimed squarely at developers rather than enterprise API customers.

The architecture is sparse MoE: 30 billion total parameters, 128 experts, 8 active per token. At inference time the effective compute load is closer to a 3B model. Minimum hardware is one H100 at FP8. Nick Frosst, Cohere co-founder, ran it on a Mac Studio via MLX at roughly 20GB RAM in the launch demo.

Key Numbers

MetricValue
Total parameters30B
Active parameters3B
Context window256K tokens
Max output64K tokens
AA Coding Index33.4
Terminal-Bench v255.1%
SWE-Bench Verified pass@1080.2%
Min hardware1× H100 @ FP8
LicenseApache 2.0

The AA Coding Index score of 33.4 puts North Mini Code ahead of Devstral Small 2 (24B dense), Nemotron 3 Super (120B-A12B), and Mistral Small 4 (119B-A6B) — all models larger by parameter count. Cohere reports 2.8x output throughput and 30% lower inter-token latency over Devstral Small 2 on identical hardware.

The Verbosity Problem

The benchmark headline masks a production cost. Artificial Analysis measured North Mini Code generating three times the output tokens of comparable models on equivalent tasks. For teams running agentic coding pipelines, verbosity compounds: higher token count means higher inference cost, higher latency, and larger context payloads passed back into subsequent calls. A model that generates clean, concise outputs at slightly lower benchmark scores can be significantly cheaper to operate at volume.

The 55.1% Terminal-Bench v2 and 80.2% pass@10 SWE-Bench Verified figures are from Cohere’s own reporting. The pass@10 SWE-Bench number measures whether at least one of ten attempts succeeds — not the resolve rate on a single pass — so it is not directly comparable to the pass@1 numbers on the main leaderboard.

License

Weights ship Apache 2.0 on Hugging Face in both BF16 and FP8 formats. One caveat: the Hugging Face model card adds a non-commercial note and an acceptable-use addendum that does not appear in Cohere’s blog post. Verify both before deploying in commercial workflows. Both standard and quantized builds are available for Ollama, LM Studio, and llama.cpp.

North Mini Code is available through the Cohere API, OpenCode (free), OpenRouter, and Hugging Face.

Context

This is Cohere’s first model after Command A+ in May, which was a 218B MoE enterprise system behind a paid API. North Mini Code goes the opposite direction: open weights, developer-first, designed to run locally. The move into the open-weights coding agent market puts Cohere alongside Mistral (Devstral), Moonshot (Kimi K2.7-Code), and Meta (Llama 4) competing for engineering teams that want capable coding agents they can self-host.

The single-H100 deployment story matters. As Fable 5 at $50/M output and managed agentic infrastructure at $0.08/hr become the cloud-native defaults, teams that can run comparable capability on hardware they already own have a real cost argument. North Mini Code’s verbosity issue narrows that argument, but the architecture is pointed in the right direction.