Cohere: North Mini Code
Open-weight agentic coding model achieving 80.2% pass@10 on SWE-Bench Verified with only 3B active parameters for efficient single-H100 deployment and a ~250K context window.
Specifications
| Attribute | Value |
|---|---|
| Lab | Cohere |
| Tags | Coding Agentic Open Weight |
| Release Date | 2026-06 |
| Context Window | 256,000 tokens |
| Input Price / 1M | $0.00 |
| Output Price / 1M | $0.00 |
| Input Modalities | Text |
| Output Modalities | Text |
Strengths
- 80.2% pass@10 on SWE-Bench Verified with vendor-reported RLVR reaching 83.2%
- Only 3B active parameters from 30B total (128 experts, 8 active) – efficient on one H100
- Free pricing for input and output per million tokens – zero cost to use
- ~250K token context window (64K max generation) for complex agentic workflows
- Open-source under Apache 2.0 license – self-host or use via OpenRouter
Weaknesses
- Weak on non-coding agentic tasks: 14% GDPval-AA and 37% τ²-Bench Telecom
- Generates ~3x more tokens than comparable models – factor this into total cost
- No multimodal support (text-only)
- Limited to code and terminal interactions – not a general-purpose model
Best For
- Agentic code generation and repair (SWE-bench tasks)
- Terminal-based coding workflows (Terminal-Bench)
- Long-context code understanding (256K tokens)
- Self-hosted development environments with single GPU
In Depth: North Mini Code
Benchmark Performance
North Mini Code achieves 80.2% pass@10 on SWE-Bench Verified, placing it in a strong agentic coding position despite ranking #74 overall (LMRank 8.0). For comparison, GPT-5.5 Pro at rank #5 overall but costs $30/$180 per M tokens.
On SWE-Bench Verified, North Mini Code reports 80.2% pass@10 (vendor-claimed). Vendor RLVR fine-tuning boosts this to 83.2% pass@1. On Terminal-Bench v2, it achieves 55.1% pass@10, rising to ~63% with RLVR. These results come from a 30B total MoE with only 3B active parameters. Independent tests show weaker non-coding scores: 14% on GDPval-AA and 37% on τ²-Bench Telecom.
Pricing & Value
North Mini Code is free: $0.00 input and $0.00 output per million tokens. Claude Fable 5, the top-ranked model at 9.9/10, costs $10.00 input and $50.00 output per M tokens.
At zero cost, North Mini Code offers a per-point value unmatched in the top 100. Compare to GPT-5.5 Pro ($30/$180 per M tokens, LMRank 9.5): that model costs $210 total per M tokens for a 9.5 score, while North Mini Code gives a 8.0 score for free. Note that Cohere reports North Mini Code generates roughly three times more tokens than typical models, so total usage costs may still accumulate in hosted settings. For self-hosting on one H100, compute cost is the only expense.
Who Should Use This
North Mini Code suits developers building autonomous coding agents on a budget. Researchers needing open-weight access for fine-tuning. Teams deploying on a single H100.
- Agentic coding developers: free pricing and strong SWE-bench results – token generation is 3x higher than peers
- Open-weight enthusiasts: Apache 2.0 license and self-hosting with vLLM – limited to coding tasks
- Budget-conscious teams: zero API cost for input/output – not suitable for general-purpose or multimodal tasks
- Anti-persona: teams needing general reasoning or multimodal support should consider GPT-5.5 Pro or Claude Opus 4.8
Release & Version History
Released June 9, 2026, North Mini Code is Cohere's first model in the North family and their first open developer model under Apache 2.0.
Version 1.0.0 launched on June 9, 2026, with 30B total parameters (128 experts, 8 active). A subsequent RLVR fine-tune (same base, version not separately versioned) improved SWE-bench pass@1 from vendor-claimed 80.2% to 83.2% and Terminal-Bench v2 from 55.1% to ~63%. OpenRouter added the model on June 17, 2026. No further updates released as of the current report.