Last updated: August 12, 2026. This is a developer-facing review of North Mini Code 1.0, Cohere's open-weight agentic coding model (30B total parameters, 3B active). Numbers below are taken from Cohere's public announcement and widely cited third-party write-ups; where a figure is not independently verified, we say so explicitly.
Hands-on verdict (what still holds after the launch pulse)
- Best fit: local / sovereign agentic coding on a single H100-class GPU (FP8), OpenCode-style terminal agents, Apache 2.0 commercial use.
- Not the pitch: absolute top of every coding leaderboard. Cohere positions it as fast + open + agentic, not "beats every dense frontier model."
- Free path is real: weights on Hugging Face under Apache 2.0; use OpenCode and local runtimes (including Ollama-style workflows where community packaging exists) without per-token billing.
- Freshness note: the June 2026 launch drove a large search spike; this refresh re-centers North Mini Code 1.0 versioning, free usage, model specs, and benchmark tables so the page matches what people still query.
What is the North Mini Code model?
North Mini Code (release id North-Mini-Code-1.0) is Cohere's first open agentic coding model in the North family. It is a sparse mixture-of-experts (MoE) model with 30B total parameters and about 3B active per token, trained for code generation, agentic software engineering, and terminal workflows.
Official launch materials (June 9, 2026) and the Hugging Face model card emphasize:
- Apache 2.0 license (commercial use permitted under the license terms)
- BF16 and FP8 weight releases on Hugging Face (CohereLabs/North-Mini-Code-1.0, FP8 variant)
- Compatibility with agent harnesses such as OpenCode, plus Cohere API / Model Vault paths for managed use
North Mini Code 1.0 model specs
| Spec | North Mini Code 1.0 | | --- | --- | | Release name | North-Mini-Code-1.0 | | Architecture | Sparse MoE Transformer (decoder-only) | | Total parameters | 30B | | Active parameters | ~3B per token | | License | Apache 2.0 | | Context length | 256K total context; up to 64K max generation (per official snapshot) | | Optimized for | Code generation, agentic SE, terminal tasks | | Weight formats | BF16, FP8 (quantized) | | Typical serving note | Cohere materials reference single H100-class GPU at FP8 for practical local serving |
HumanEval and some older code-eval suites: official data not verified for a single canonical North Mini Code 1.0 HumanEval score in the sources checked for this update—prefer the agentic benches Cohere actually published against.
North Mini Code benchmark results
Cohere evaluates agentic coding primarily with SWE-Bench Verified, SWE-Bench Pro, Terminal-Bench v2, and Terminal-Bench Hard (plus coding-index style aggregates).
Cohere-reported agentic scores (North Mini Code 1.0)
| Benchmark | North Mini Code 1.0 (reported) | Notes | | --- | --- | --- | | SWE-bench Verified | 67.6% | Widely repeated from Cohere launch materials / third-party summaries | | SWE-bench Pro | 40.2 | Harder / newer issue set | | Terminal-Bench v2 | 36.0 | Terminal / shell agent tasks | | Terminal-Bench Hard | 31.1 | Harder terminal suite | | Artificial Analysis Coding Index | 33.4 | Cohere cites this competitive mid-size coding index score |
Independent lab-to-lab numbers can differ by harness, temperature, and scaffold. If you need production selection data, re-run on your repos and agent harness—not only leaderboard rows.
Comparison notes (Qwen / Devstral class; DeepSeek)
Public launch write-ups compare North Mini Code mainly against similar-size open coders (e.g. Qwen 3.x MoE mid-size variants, Devstral Small 2, Gemma-class MoEs). Directionally:
| Priority | Lean toward | | --- | --- | | Permissive Apache 2.0 + sovereign local deploy | North Mini Code 1.0 | | Peak raw coding index among same-size opens (as of mid-2026 commentary) | Often Qwen mid-size MoE variants on some indexes—verify current Artificial Analysis / vendor tables | | Battle-tested agentic coding product history | Devstral class models still common baselines | | DeepSeek-Coder head-to-head on the same harness | Official data not verified for a single locked table in this refresh—do not invent scores |
Sebastian Raschka's June 2026 write-up notes Qwen3.6-35B-A3B ahead of North Mini Code on the four agentic rows Cohere published (example: SWE-bench Verified ~73.4 vs 67.6 in that summary). Treat cross-blog tables as secondary to primary Cohere/HF cards.
How to use North Mini Code for free
This section is for queries like north mini code free / local free access.
1) Free weights (no API bill)
- Download North-Mini-Code-1.0 (BF16 or FP8) from Hugging Face under Apache 2.0.
- Serve with your preferred inference stack (vLLM, TGI, llama.cpp-style paths if/when community GGUF packaging exists, etc.).
- Point your coding agent at the local OpenAI-compatible endpoint.
You pay for hardware/electricity, not Cohere tokens.
2) Free / open agent path with OpenCode
Cohere explicitly trained and demoed OpenCode integration. Typical free path:
- Install OpenCode (or your fork of the open coding agent).
- Configure the model provider to either:
- a local OpenAI-compatible server running North Mini Code 1.0, or
- a provider that exposes the open weights without forcing a paid Cohere plan (community routers vary—check current docs).
- Run repo tasks (edit / test / terminal) and measure pass rate on your issues.
3) Ollama-style local workflow
Community packaging moves quickly. Practical pattern:
- Prefer official HF weights first.
- If an Ollama model tag for North Mini Code 1.0 is published by Cohere or a trusted quant maintainer, pull it with ollama pull <tag> and set your agent model to that tag.
- If no official Ollama tag exists yet, serve HF FP8/BF16 yourself and connect OpenCode / Continue / Aider-style tools to localhost.
Do not assume a specific Ollama tag name forever—tags churn. The durable free story is Apache 2.0 weights + local server + OpenCode.
4) Paid convenience (optional)
Cohere API / Model Vault / OpenRouter-style hosts can be faster to try before you buy GPUs. That is convenience, not a requirement for free use of the open model.
Changelog / version notes (North Mini Code 1.0)
| Version | Date | What changed | | --- | --- | --- | | North Mini Code 1.0 (North-Mini-Code-1.0) | 2026-06-09 | Initial public open-weight release: 30B-A3B MoE, Apache 2.0, HF BF16/FP8, OpenCode-oriented agentic coding focus | | Post-1.0 patches | — | No separate minor version line verified for this page refresh; watch Cohere blog + HF model card for successors |
Search variants like north-mini-code-1.0 map to this same release id.
How to get started (practical checklist)
- Skim the official model card on Hugging Face for license, context, and serving notes.
- Smoke-test via API or a hosted router with 2–3 tasks from your real repo.
- Switch to local FP8 if latency/$ and data residency matter.
- Bind OpenCode (or mini-SWE-agent / SWE-agent) and measure task completion, not vibes.
- Track tokens/sec and cost per merged PR—North Mini Code's design bet is throughput + open license, not maximum intelligence index.
Frequently asked questions
What is North Mini Code 1.0?
North Mini Code 1.0 is Cohere's open agentic coding model: ~30B total / ~3B active MoE parameters, Apache 2.0, optimized for code and terminal agents, released June 9, 2026.
Is North Mini Code free?
The weights are free to download under Apache 2.0. Local inference is free of per-token fees (you supply compute). Hosted APIs may charge. That is the accurate "north mini code free" answer.
How good is North Mini Code on SWE-bench?
Cohere-reported SWE-bench Verified ~67.6% and SWE-bench Pro ~40.2 for the 1.0 launch table. Replicate on your harness before production selection.
North Mini Code vs Qwen-Coder / DeepSeek-Coder?
Vs mid-size Qwen MoE coders: Qwen often leads some coding indexes; North Mini Code competes on open license + agentic packaging + speed narrative. Vs DeepSeek-Coder: official side-by-side data not verified here—compare on identical SWE-agent settings instead of marketing slides.
Can I fine-tune North Mini Code?
Apache 2.0 generally allows modification and commercial distribution under the license terms—confirm against the license text and any HF card notices for your jurisdiction and product.
Review conclusion (updated August 2026)
North Mini Code 1.0 remains a strong pick when you want an open coding model you can actually own: Apache 2.0, 30B/3B MoE, 256K-class context, and a free local path through HF + OpenCode. It is not automatically the #1 scoreboard model in its class on every index, and that is fine—the product thesis is agentic coding you can self-host.
If your constraint is maximum closed-API quality regardless of license, use a frontier API. If your constraint is sovereign, free-to-run coding agents, North Mini Code 1.0 still belongs on the shortlist—re-benchmark on your harness after each competitor drop.
Sources checked for this refresh include Cohere/HF launch materials (North Mini Code 1.0), Artificial Analysis coding-index citations in those materials, and independent technical summaries (e.g. agentic bench tables). Always re-verify live model cards before shipping.