推出 North Mini Code：Cohere 為開發者打造的首款模型

Back to Articles Introducing North Mini Code: Cohere’s First Model For Developers Enterprise Article Published June 9, 2026 Upvote 32 +26 Cohere Code Agents Team coherecode Follow CohereLabs All co-authors listed below Today, we are releasing North Mini Code, a 30B-parameter Mixture-of-Experts model with 3B active parameters with powerful agentic coding capabilities, available on Hugging Face under the Apache 2.0 license. North Mini Code is the first model in Cohere’s new family of models, and is specifically designed and trained for agentic software engineering tasks. Figure 1: North Mini Code’s performance in agentic coding tasks and complex code generation benchmarks, compared to leading open-source models of similar size. See here for the details of our benchmarking methodology. North Mini Code is optimized for complex software engineering workflows, terminal-based agentic tasks, and high-quality code generation. On Artificial Analysis’ Coding Index, North Mini Code achieves a score of 33.4, outperforming Qwen3.5 (35B-A3B), Gemma 4 (26B-A4B), Devstral Small 2 (24B Dense), and even substantially larger models such as Nemotron 3 Super (120B-A12B), Mistral Small 4 (119B-A6B), and Devstral 2 (123B).1 It ranks among the strongest open-source coding models in its size class. Try North Mini Code in OpenCode Real-world code agents depend on model quality and robustness across agent harnesses. We trained North Mini Code using multiple scaffolds rather than optimizing for a single one. This approach enables North Mini Code to serve as a reliable foundation for coding agents such as OpenCode. Architecture Figure 2: North Mini Code is a Mixture-of-Experts Transformer decoder with interleaved sliding-window self-attention and full self-attention. North Mini Code is a decoder-only Transformer-based sparse Mixture-of-Experts model. It uses our efficient attention implementation, interleaved between sliding-window attention with RoPE and global attention with no positional embeddings, in a 3:1 ratio [1]. The feed-forward block is an MoE block with 128 experts, of which 8 are activated per token. Each expert block is an FFN block with SwiGLU activation. The router applies a sigmoid activation function to the logits before the top-k selection. We also use a single dense layer before the sparse layers. Post-Training for Coding Excellence Figure 3: The post-training pipeline is made up of two phases of supervised fine-tuning (SFT) and a phase of agentic reinforcement learning with verifiable rewards (RLVR) targeting software engineering and terminal tasks. We post-train North Mini Code using a two-stage cascaded supervised fine-tuning (SFT) followed by reinforcement learning with verifiable rewards (RLVR), focusing on agentic coding. Our first stage SFT data focuses on coding capabilities that are integrated within a wider mix for robustness and usability. The datamix includes programming, reasoning, and instruction following across a large variety of domains where the code datasets correspond to 70% of trainable tokens, 43% agentic tool-use data, and 27% single-turn competitive or scientific programming data. In the second stage SFT, we use a 4.5 billion token data mixture from only agentic and reasoning-driven samples, where code data forms 61% of trainable tokens. This mixture comprises our highest-quality data across coding and wider agentic tasks where tool calls and completions are verified as executable and correct. Our internal data pipeline heavily relies on containerised agentic coding environments. We maintain a disjoint subset of these environments for use in synthetic SFT data generation and RLVR. The majority are based on software engineering tasks from real-world repositories, while the rest are terminal-based agentic tasks sourced from open-source and internal datasets. In total, we used over 70k verifiable tasks across ~5k unique repositories. We deduplicate our environments against the repository sources from SWE-Bench [2] and SWE-Bench-Pro [3] to avoid source leakage during evaluation [4]. We used 64K and 128K context lengths for the first and second stages of SFT, respectively. This “long-to-longer” cascade approach (similar to [5, 6]) enables bipartite training on valuable shorter data, establishing a robust performance baseline, followed by targeted long-context training only on high-quality verified samples. Without multi-stage training, the 20B non-code tokens during the initial training stage often dominated the 1.5B tokens of high-quality code data in later training, producing poorer performance and higher behavioral conflicts from data trends differing between stages. Anecdotally, training on a near-complete length distribution of samples produced shorter final trajectories during evaluation than training on a truncated distribution up to 64K only. Instead of optimising North Mini Code towards quantitative metrics during SFT, we adopted an approach strictly using SFT as priming for RLVR. The data mixture optimises sampling diversity and pass@K (for high K) in downstream stages. We use sample-level filtering to remove any pathologies such as invalid tool calls, erroneous whitespace generation, malformed special tokens, or hallucinated citations. Artifacts or hyperparameters producing undesirable RLVR behaviours (e.g., low entropy, invalid structured generations) were pruned via ablations. The final SFT model achieves 80.2% pass@10 on SWE-Bench Verified [2] and 55.1% pass@10 on Terminal-Bench v2 [7]. Robustness Across Harnesses Harness robustness improves model usability in realistic software development settings, where agents encounter diverse and unpredictable tooling environments. These environments differ not just in prompting but in fundamental tool-use modality, For instance, SWE-Agent [8] exposes a relatively rich agent-CLI interface with specialized commands (bash, str_replace_editor and submit tools) and templated observations; mini-SWE-agent [9] strips this down to a single bash tool, with raw stdout from shell as the only feedback; and OpenCode [10] uses fine-grained individually typed tools (edit, grep, todowrite and task etc) returning structured JSON responses. Figure 4: To power a variety of agentic coding harnesses, North Mini Code is exposed to a variety of coding harnesses during the second SFT stage. We address cross-harness generalization by introducing a small amount of additional benchmark harness data (6% of the SFT mix, compared to 50% of the chosen SWE-Agent harness) during the second SFT stage. Specifically, this data mix yields a 10% gain on the evaluation with OpenCode harness while maintaining performance with SWE-Agent on SWE-Bench Verified, demonstrating that cross-harness transfer can be cheaply acquired without degrading benchmark performance. Notably, North-Code-Mini achieves 61.0% pass@1 using mini-SWE-Agent, where the improvement emerged for free in the cross-task, cross-harness settings, suggesting that harnesses with overlapping tool capabilities share enough representational structure for positive transfer. We also observe minimal data conflict when training on hybrid harness data, indicating that skills required by different harnesses are usually complementary rather than contradictory. Similarly, the official Terminal-Bench uses its own Terminus 2 harness, where all the agent-CLI interactions are communicated via plain-text chat turns (instead of native tool calling). In order to prime our models on Terminus 2, we include a small amount of data (less than 20%) in a plain-text format in the data mixture, which has proved sufficient for the model to naturally generalise across. Interestingly, we also find that it’s crucial to introduce sufficient variations in the various harnesses (akin to data augmentation) in order to force the model to properly establish the link between instructions and behaviours rather than simply regurgitating a fixed template without understanding, and this is especially impo