Cohere Releases North Mini Code: A 30B Agentic Coding Model Optimized for Single-GPU Inference

Cohere released North Mini Code on June 9, 2026, positioning the system as a 30 billion parameter agentic coding agent. [3] The model operates as a sparse mixture-of-experts Transformer featuring 128 total experts, with eight activated per token. [3] This specific architecture gives the system the effective compute profile of a 3 billion parameter dense model during inference, despite its 30 billion total parameters. [3] Cohere co-founder Nick Frosst demonstrated the model running locally on a Mac Studio utilizing MLX with approximately 20 gigabytes of RAM. [3]

Cohere made the model freely available to developers under an Apache 2.0 license. [9] The release directly targets developers seeking open-source coding agents that can operate efficiently on a single NVIDIA H100 GPU. [3] North Mini Code supports a 256,000-token total context window alongside a 64,000-token maximum generation length. [9] The model weights can be downloaded from Hugging Face in BF16, FP8, and W4A16 formats. [9]

Independent testing conducted by Artificial Analysis ranks North Mini Code eighth out of 127 comparable open-weight models on output speed. [3] The model achieved an output rate of 210 tokens per second during these third-party tests. [3] Furthermore, Artificial Analysis recorded a time to first token of 0.25 seconds for North Mini Code, establishing a notable speed advantage compared to the class median of 1.95 seconds. [3]

The system achieved a score of 33.4 on the Artificial Analysis Coding Index. [9] In addition to the downloadable weights, Cohere made the model accessible via its own API, the Cohere Model Vault, and OpenRouter. [9] For inference hardware, the minimum requirement to run the system is one H100 at either FP8 or FP4. [9]

Unlike standard AI models that fine-tune a general-purpose base model on code data, Cohere built North Mini Code from scratch specifically for agentic software engineering tasks. [3] The system is highly optimized for multi-file code review, architecture mapping, sub-agent orchestration, and terminal tasks. [3] It is also integrated into existing scaffolds including OpenCode, SWE-Agent, and mini-SWE-Agent. [3]

However, the system exhibits notable verbosity during operational output. [3] During the Artificial Analysis benchmarking process, the model generated 75 million output tokens. [3] This volume represents three times the output of comparable models, which achieved a class median of 25 million output tokens in the same tests. [3]

Cohere has applied its AI agent platforms to internal operations by connecting the system to the cloud security platform Wiz through a custom Model Context Protocol (MCP) server. [7] The resulting security agent automates incident response workflows, handling tasks from triaging critical findings to drafting reports and updating Wiz status. [7] Prior to implementing this agentic system, manually investigating an affected asset and searching for existing tracking tickets could take between 30 minutes to two hours per critical finding. [7]

In a separate engineering application, Cohere utilized AI coding agents to automate the feedback loop required for maintaining a long-lived software fork. [8] The company applied this method to its internal fork of vLLM after a routine upstream release broke its cohere-transcribe-03-2026 ASR model. [8] The agentic workflow compressed the time required to absorb a new upstream release from weeks of intermittent developer attention to just days of mostly unattended agent time. [8]

The release of North Mini Code introduces the clearest open-source contender to emerge in 2026 for developers deciding whether to route agentic coding workloads through a self-hosted open-source model or a managed API. [3] Its time to first token of 0.25 seconds significantly undercuts the 1.95-second class median of comparable open-weight models evaluated by Artificial Analysis. [3] By utilizing a 128-expert architecture that activates only eight experts per token, Cohere matches the inference compute profile of a 3 billion parameter dense model. [3] What the sources don't address: How the model's generation of 75 million output tokens impacts downstream hosting costs compared to the 25 million token class median.