Cohere Launches North Mini Code for Developers
๐กCohere's first dedicated coding model could be a game-changer for your IDE-integrated AI workflows.
โก 30-Second TL;DR
What Changed
First Cohere model specifically tailored for coding tasks
Why It Matters
This model provides developers with a specialized alternative to general-purpose LLMs, potentially improving code generation accuracy and latency for coding assistants.
What To Do Next
Evaluate North Mini Code against your current coding assistant model using the Cohere API to compare latency and code quality.
Key Points
- โขFirst Cohere model specifically tailored for coding tasks
- โขOptimized for developer-centric workflows and integration
- โขExpands Cohere's enterprise-grade model ecosystem
๐ง Deep Insight
Web-grounded analysis with 11 cited sources.
๐ Enhanced Key Takeaways
- โขNorth Mini Code is a 30-billion-parameter Mixture-of-Experts (MoE) model with only 3 billion active parameters, designed to reduce inference costs while maintaining strong performance for agentic coding tasks.
- โขThe model is open-sourced under the Apache 2.0 license, making it freely available on Hugging Face for developers to modify, deploy, and commercialize without licensing restrictions.
- โขIt is specifically optimized for complex software engineering workflows, terminal-based agentic tasks, and high-quality code generation, achieving a score of 33.4 on Artificial Analysis' Coding Index.
- โขNorth Mini Code supports a substantial context length of 256K tokens and can generate outputs up to 64K tokens, enabling it to handle entire mid-sized codebases and produce significant blocks of code in a single pass.
- โขThe model was post-trained using a two-stage cascaded supervised fine-tuning (SFT) followed by reinforcement learning with verifiable rewards (RLVR), with 70% of trainable tokens dedicated to coding datasets.
๐ Competitor Analysisโธ Show
| Model Name | Type | Total Parameters (Active) | Context Window | Key Benchmarks (Artificial Analysis Coding Index / SWE-Bench Verified) | Pricing (per 1M tokens) |
|---|---|---|---|---|---|
| Cohere North Mini Code | Open-source (Apache 2.0) | 30B (3B) | 256K input, 64K output | 33.4 (Coding Index) / N/A (SWE-Bench Verified) | Free (model weights) |
| Claude Opus 4.6/4.7 | Proprietary | N/A | 1M | 80.8% (SWE-Bench Verified) | $5.00 input / $25.00 output |
| GPT-5.3 Codex | Proprietary | N/A | N/A | 56.8% (SWE-Bench Pro) | N/A |
| Gemini 3.1 Pro | Proprietary | N/A | 1M | 80.6% (SWE-Bench Verified) | $1.25 input / $10 output |
| Kimi K2.5 / Kimi-Dev-72B | Open-source | 1T (K2.5) / 72B (Kimi-Dev-72B) | 1M (K2.5) / 131K (Kimi-Dev-72B) | 85% (LiveCodeBench K2.5) / 60.4% (SWE-Bench Verified Kimi-Dev-72B) | Open-source |
| Qwen3.5 (35B-A3B) | Open-source | 35B (3B) | N/A | Outperformed by North Mini Code (Coding Index) | Open-source |
| DeepSeek V3.2 | Open-source | 685B | N/A | Strong code generation | $0.28 input / N/A output |
| Nemotron 3 Ultra | Open-source | 253B | N/A | Outperformed by North Mini Code (Coding Index) | Open-source |
| Mistral Small 4 | Open-source | 119B (6.5B) | N/A | Outperformed by North Mini Code (Coding Index) | Open-source |
๐ ๏ธ Technical Deep Dive
- Model Architecture: Decoder-only Transformer-based sparse Mixture-of-Experts (MoE) model.
- Parameter Count: 30 billion total parameters with 3 billion active parameters during inference.
- Attention Mechanism: Uses an efficient attention implementation, interleaved between sliding-window attention with RoPE (Rotary Positional Embeddings) and global attention with no positional embeddings, in a 3:1 ratio.
- Expert Blocks: The feed-forward block is an MoE block with 128 experts, of which 8 are activated per token. Each expert block uses SwiGLU activation.
- Context Length: Supports a 256K token input context window and a 64K token maximum output length.
- Training Methodology: Post-trained using a two-stage cascaded supervised fine-tuning (SFT) followed by reinforcement learning with verifiable rewards (RLVR), specifically focusing on agentic coding.
- Training Data Mix: The SFT data includes programming, reasoning, and instruction following across various domains, with code datasets corresponding to 70% of trainable tokens, 43% agentic tool-use data, and 27% single-turn competitive or scientific programming data.
- Tool Use Capabilities: Specifically trained with tool-use capabilities for agentic coding, supporting tool descriptions using JSON schema and designed for interleaved thinking in agentic steps.
- License: Released under the Apache 2.0 license.
๐ฎ Future ImplicationsAI analysis grounded in cited sources
โณ Timeline
๐ Sources (11)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
Weekly AI Recap
Read this week's curated digest of top AI events โ
๐Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: Hugging Face Blog โ
