DEV.co
Open-Source LLM · CohereLabs

North-Mini-Code-1.0

North Mini Code is a 30B-parameter sparse mixture-of-experts model from Cohere Labs optimized for code generation and agentic software engineering tasks. It uses 3B active parameters per token, supports 256K context length with 64K max output, and is licensed under Apache 2.0. The model is designed to work with tool-use (bash, file operations) and interleaved reasoning for autonomous coding workflows.

Source: HuggingFace — huggingface.co/CohereLabs/North-Mini-Code-1.0
30.5B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
37.2k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
DeveloperCohereLabs
Parameters30.5B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads37.2k
Likes514
Last updated2026-06-15
SourceCohereLabs/North-Mini-Code-1.0

What North-Mini-Code-1.0 is

Decoder-only Transformer with sparse MoE architecture: 128 experts with 8 activated per token. Hybrid attention: sliding-window (RoPE) and global attention in 3:1 ratio. SwiGLU activation in FFN blocks. Trained via two-stage cascaded SFT + RLVR (reinforcement learning with verifiable rewards) focused on agentic coding. Supports tool-use via JSON schema function definitions and interleaved thinking (requires vLLM main branch or Transformers from source). Native chat template support in Transformers.

Quickstart

Run North-Mini-Code-1.0 locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="CohereLabs/North-Mini-Code-1.0")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Autonomous code agents and software engineering tasks

Explicitly trained for agentic workflows. Supports tool-use (bash, file I/O) and interleaved reasoning. Benchmarked on SWE-Bench Verified/Pro and Terminal-Bench; designed for multi-step, verifiable coding tasks.

Long-context code analysis and generation

256K context window enables analysis of large codebases, complex documentation, or multi-file projects. 64K output length supports generating substantial code artifacts in a single response.

Efficient inference in resource-constrained environments

Sparse MoE with only 3B active parameters per token provides effective model capacity (30B total) while reducing compute footprint compared to dense 30B alternatives. Suitable for cost-sensitive or edge deployments.

Running & fine-tuning it

ESTIMATE: 60–120 GB VRAM for fp16 inference of 30B sparse model (fp32: ~120–180 GB). Actual footprint depends on sparse expert routing overhead and attention implementation. Model card recommends tensor parallelism (`-tp 2` in vLLM for single node or multi-GPU setups). Batch inference and long sequences (up to 256K tokens) require substantial VRAM. Single-GPU inference is impractical; dual-GPU minimum recommended. Exact VRAM/precision breakdown not stated in card—verify with benchmark on target hardware.

Not documented in model card. No mention of LoRA, QLoRA, or fine-tuning recipes. Model was trained via SFT + RLVR; no guidance on whether these techniques apply to downstream fine-tuning or if the sparse MoE architecture supports parameter-efficient adaptation. Requires direct contact with Cohere Labs or empirical testing.

When to avoid it — and what to weigh

  • Strict real-time latency requirements without GPU clustering — 30B sparse model requires significant VRAM and compute. Single-GPU inference will be slow. Requires tensor parallelism (model card recommends `-tp 2` in vLLM) for acceptable latency.
  • Non-code generative tasks requiring general-purpose LLM reasoning — Model is specialized for code and agentic engineering. No evidence from card that it matches general-purpose models (GPT, Llama) on text generation, summarization, or non-technical reasoning tasks.
  • Requirement for production-grade security and compliance audits — No mention of security audits, red-teaming results, or vulnerability disclosures in card. Agentic code execution (bash, file operations) introduces risk if tool outputs are not carefully validated by the operator.
  • Offline/edge deployment without vLLM or custom inference pipeline — Requires vLLM main branch (or Transformers from source) and Cohere's melody library for tool-call parsing. Standard llama.cpp, Ollama, or TGI support is not documented.

License & commercial use

Apache 2.0 license. Permissive OSI-approved license allowing use, modification, and distribution.

Apache 2.0 is a permissive, business-friendly license that permits commercial use, derivative works, and proprietary applications. No restrictions stated or apparent. However, verify with legal counsel if deploying agentic code execution (bash, file operations) in production to ensure liability and safety compliance are acceptable for your use case.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationStrong
License clarityClear
Deployment complexityModerate
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

Model is designed for tool-use (bash execution, file I/O) in agentic workflows. No security audit, red-teaming results, or adversarial robustness evaluation documented. Agentic code execution poses risk if tool outputs are not validated or if the model is manipulated to execute unintended commands. Recommend sandboxing tool execution, rate-limiting, and input validation. No mention of jailbreak resistance or safety alignment guarantees. Private deployment (not exposed to untrusted users) reduces surface area.

Alternatives to consider

Llama 3.1 405B or 70B (Meta)

Larger general-purpose LLM with strong code and reasoning capabilities. Better for non-agentic coding and general-domain tasks. Higher VRAM cost but no sparse expert overhead; broader fine-tuning community.

DeepSeek Coder or Qwen QwQ (Alibaba/DeepSeek)

Alternative code-specialized models with large context windows. Unknown if they support tool-use as natively as North Mini Code. May offer different cost/performance trade-offs.

Claude or GPT-4 (Anthropic/OpenAI)

Proprietary, hosted agentic LLMs with extensive tool-use support, safety training, and SLAs. Higher per-token cost but no infrastructure overhead. Better for teams avoiding self-hosting.

Software development agency

Ship North-Mini-Code-1.0 with senior software developers

Start with vLLM (recommended) or Transformers pipeline. Review hardware requirements (dual-GPU, 60–120 GB VRAM) and validate tool-use safety for your deployment. Contact Cohere Labs for production support and fine-tuning guidance.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

North-Mini-Code-1.0 FAQ

Can I use North Mini Code commercially?
Yes. Apache 2.0 is a permissive license that permits commercial use, modification, and distribution without royalty. However, consult legal counsel if deploying agentic code execution (bash, file operations) in production to ensure liability, safety, and compliance align with your use case.
What hardware do I need to run this model?
Estimate 60–120 GB VRAM for fp16 inference. Dual-GPU minimum recommended; tensor parallelism (`-tp 2`) is standard. Single-GPU inference is impractical. Exact VRAM depends on sparse expert routing and attention overhead—verify on your target hardware. vLLM with main branch is recommended for serving.
Does North Mini Code support fine-tuning?
Not documented in the model card. No guidance on LoRA, QLoRA, or downstream fine-tuning. The model was trained via SFT + RLVR, but applicability to parameter-efficient adaptation of the sparse MoE architecture is unknown. Contact Cohere Labs or test empirically.
Can I run this model locally without vLLM?
Yes, via Transformers pipeline (standard library). However, vLLM main branch is recommended for tool-call and reasoning parsing via Cohere's melody library. Offline serving without vLLM or Transformers from source is not documented.

Work with a software development agency

DEV.co is a software development agency delivering custom software development services to companies building on open source. Our software developers and web developers design, integrate, and ship production systems — spanning web development, APIs, AI, data, and cloud. If North-Mini-Code-1.0 is part of your open-source llms roadmap, our team can implement, customize, migrate, and maintain it.

Ready to Deploy North Mini Code?

Start with vLLM (recommended) or Transformers pipeline. Review hardware requirements (dual-GPU, 60–120 GB VRAM) and validate tool-use safety for your deployment. Contact Cohere Labs for production support and fine-tuning guidance.