Private coding agents
Run a coding model in controlled infrastructure when source code cannot be sent to a shared external service.
Independent tool overview
North Mini Code at a glance
North Mini Code is Cohere's Apache-2.0 agentic coding model: a 30-billion-parameter mixture-of-experts system with 3 billion active parameters, a 256K context window, and local, API, or dedicated deployment options.
Visit the official North Mini Code site ↗
- Model ID
- north-mini-code-1-0
- Architecture
- 30B total / 3B active MoE
- Context window
- 256,000 tokens
- Maximum generation
- 64,000 tokens
- License
- Apache 2.0
- Published formats
- BF16, FP8, and W4A16
- Deployment
- Self-hosted, Cohere API, or Model Vault
Overview
What North Mini Code is
North Mini Code is Cohere's first model built specifically for agentic software engineering. It is intended to work inside coding harnesses that can inspect repositories, call terminal tools, edit files, run tests, and continue across multiple steps.
The model uses a mixture-of-experts architecture with 30 billion total parameters and about 3 billion active for each token. That lowers inference work compared with a similarly sized dense model, but it does not turn the full checkpoint into a 3B download: the official BF16 repository is roughly 61 GB.
Cohere publishes the weights under Apache 2.0 in BF16, FP8, and W4A16 formats. Teams can self-host with frameworks including Transformers, vLLM, SGLang, and Docker Model Runner, call model ID north-mini-code-1-0 through Cohere, or use a dedicated Model Vault deployment.
The model offers 256,000 tokens of context and up to 64,000 generated tokens. Cohere trained it across multiple agent harnesses and specifically highlights OpenCode compatibility, terminal tasks, repository changes, architecture mapping, code review, scientific coding, and subagent orchestration.
North Mini Code is attractive when data location, open licensing, controllable infrastructure, and operating cost matter more than having the strongest proprietary frontier model. It still needs a capable harness, safe permissions, evaluation on the team's repositories, and human ownership of shipped changes.
Use cases
Who North Mini Code is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Cost-aware agent loops
Use a small active-parameter footprint for workflows that make many model and terminal calls.
Repository maintenance
Pair the model with a harness for scoped code changes, tests, refactors, reviews, and issue resolution.
Terminal automation
Drive a narrow set of shell tools through a multi-turn ReAct-style loop with approvals and validation.
Sovereign deployment
Keep model weights, code, logs, and inference in an organization's cloud, VPC, on-premises environment, or dedicated vault.
Open-model evaluation
Benchmark a permissively licensed coding model against proprietary APIs on the team's own languages and repositories.
Capabilities
Core North Mini Code features
Agentic coding specialization
Trained for repository-level changes, terminal work, code generation, review, architecture mapping, and multi-step software tasks.
Mixture-of-experts efficiency
Uses about 3B active parameters from a 30B total model, reducing per-token computation while retaining broader parameter capacity.
Long context
Accepts up to 256K tokens and can generate up to 64K tokens, supporting substantial code and tool history when context is curated.
Apache 2.0 weights
Allows broad commercial use, modification, and redistribution subject to the license's notice and attribution requirements.
Multiple precision formats
Official BF16, FP8, and W4A16 releases provide options for quality, memory use, and serving throughput.
Common serving frameworks
The official repository documents Transformers, vLLM, SGLang, Docker Model Runner, and OpenAI-compatible local endpoints.
Cohere API access
Supports Cohere Chat V2, Chat V1, and Chat Completions through model ID north-mini-code-1-0.
Dedicated Model Vault
Cohere can operate the model in an isolated, single-tenant managed environment for production workloads.
Structured agent capabilities
Cohere lists reasoning, tool use, structured outputs, multilingual use, image inputs, safety modes, and citations for the hosted model.
Process
How the North Mini Code workflow works
Step 1
Choose the deployment boundary
Decide between a quick Cohere API evaluation, self-hosting for maximum control, or Model Vault for dedicated managed infrastructure.
Step 2
Select the model format
Benchmark BF16, FP8, or W4A16 on the target accelerators; do not assume the 3B active count means the full 30B checkpoint fits like a dense 3B model.
Step 3
Use a coding harness
Connect the model to OpenCode or another agent loop with explicit file, search, patch, terminal, and test tools.
Step 4
Load repository rules
Provide architecture notes, coding conventions, test commands, acceptance criteria, and protected areas before allowing edits.
Step 5
Restrict permissions
Use a sandbox, least-privilege credentials, network allowlists, resource limits, and approval gates for consequential actions.
Step 6
Evaluate representative tasks
Measure correct patches, test pass rate, regressions, retries, token use, latency, GPU utilization, and human review time on real repositories.
Step 7
Review every result
Inspect the diff and run appropriate tests, security checks, builds, and runtime validation before merging or deploying.
Step 8
Operate the service
Monitor capacity, queueing, context length, cost, failure modes, model updates, audit logs, and abuse or prompt-injection attempts.
Cost
North Mini Code pricing and free plan
The North Mini Code weights have no software-license fee under Apache 2.0, but self-hosting requires compute and operations. Cohere currently allows hosted API use without a token charge until rate limits are reached. Dedicated production use through Standard Model Vault is listed at $57.50 per XL instance-hour, with access and commitment pricing handled through Cohere.
Open weights
$0 software license
Download and self-host the Apache-2.0 weights in an available precision format.
- Compute, storage, bandwidth, engineering, and monitoring are separate
- The BF16 repository is roughly 61 GB
- FP8 and W4A16 can lower memory requirements
- Review the Apache 2.0 notice and redistribution obligations
Cohere hosted API
$0 until rate limits
Cohere's model page says trial and production keys can call North Mini Code without a model charge until applicable limits are reached.
- Trial rate limit is 20 requests per minute
- Published production rate limit is 500 requests per minute
- Key-level monthly caps and policies may also apply
- A free hosted endpoint should not be treated as a guaranteed long-term production price
Standard Model Vault
$57.50 per XL instance-hour
Dedicated Cohere-managed inference for organizations that want isolation without operating the serving stack.
- Generative-model access may require a waitlist
- Fixed and Flex capacity models are available
- Monthly and annual commitment rates require Cohere
- Encrypted Vault pricing is separate
Pricing checked . Check current pricing at the source ↗
Assessment
North Mini Code strengths and limitations
Where it stands out
- Permissive Apache 2.0 license supports commercial self-hosting and modification
- Small active-parameter footprint is designed for efficient agentic workloads
- 256K context and 64K output support long repository and tool-use sessions
- Official precision variants and common serving-framework examples reduce deployment friction
- Supports self-hosted, shared API, and dedicated managed deployment models
- Training across multiple harnesses reduces dependence on one proprietary agent shell
- Cohere publishes model specs, rate limits, license, weights, and dedicated-inference pricing
What to consider
- North Mini Code is a model component, not a complete IDE, coding agent, sandbox, permission system, or deployment workflow
- The full 30B checkpoint must be stored and served even though only about 3B parameters are active for each token
- Local deployment still requires suitable accelerators, memory planning, serving software, observability, and model operations
- Quantization can change quality, speed, memory use, and tool-call reliability; benchmark the exact published format
- A 256K window does not ensure accurate use of every file or token, so targeted retrieval and context management remain important
- Cohere's throughput, latency, and benchmark comparisons are vendor-run and may not reproduce on different hardware, harnesses, prompts, or repositories
- The hosted API's current free pricing is bounded by rate and key limits and may not provide the economics or service guarantees a production workload needs
- Generated code can contain security flaws, regressions, invented APIs, weak tests, or incomplete fixes
- Terminal access creates prompt-injection, secret-exposure, data-loss, and supply-chain risk unless tools are tightly sandboxed
- Teams must still review licenses for generated code, dependencies, training-data concerns, and any redistributed model modifications
Compare
North Mini Code alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Coding
Devstral 2
Choose Devstral 2 when comparing Mistral's newer open coding-model family and its Vibe agent workflow.
Explore Devstral 2 →Coding
Qwen3-Coder-Next
Choose Qwen3-Coder-Next for another open-weight, efficiency-focused model designed for agentic coding.
Explore Qwen3-Coder-Next →Coding
GPT-5.3-Codex
Choose GPT-5.3-Codex when frontier hosted capability and a managed OpenAI agent ecosystem matter more than weight access.
Explore GPT-5.3-Codex →Coding
Gemini CLI
Choose Gemini CLI when you want an open-source terminal harness with Google-hosted models and a low-friction entry path.
Explore Gemini CLI →Questions
North Mini Code FAQs
What is North Mini Code?
North Mini Code is Cohere's 30B-total, 3B-active mixture-of-experts model for agentic coding, terminal work, repository changes, code review, and related software tasks.
Is North Mini Code open source?
Cohere calls it open source and publishes the model weights under Apache 2.0. A precise description is an openly downloadable, permissively licensed model that can be self-hosted and modified.
Can North Mini Code run locally?
Yes. Cohere publishes BF16, FP8, and W4A16 variants and documents several local serving frameworks. Hardware needs depend on precision, context length, concurrency, cache size, and performance targets.
Does 3B active mean it needs the same memory as a 3B model?
No. About 3B parameters participate in each token's computation, but the serving system still needs access to the full 30B mixture-of-experts checkpoint. The official BF16 repository is roughly 61 GB.
How much does North Mini Code cost?
The weights have no software fee under Apache 2.0. Cohere currently provides hosted API calls without a token price until limits apply. Standard Model Vault lists North Mini Code at $57.50 per XL instance-hour. Self-hosting infrastructure is separate.
What context window does North Mini Code have?
Cohere documents a 256,000-token context window and a 64,000-token maximum generation length.
Can I use North Mini Code commercially?
The Apache 2.0 license permits commercial use, subject to its terms. Organizations should still review model notices, output and dependency licenses, data policies, and their intended distribution.
Does North Mini Code work with OpenCode?
Yes. Cohere says it was specifically trained for OpenCode compatibility, while also targeting broader coding harnesses and terminal agents.
Is North Mini Code safe to give terminal access?
Not without controls. Use a sandbox, minimal credentials, network restrictions, approved commands, resource limits, audit logs, human confirmation for consequential actions, and independent testing of every change.
Bottom line
Our North Mini Code verdict
North Mini Code is a compelling option for teams that want a permissively licensed coding model with a long context window, efficient MoE inference, and several deployment paths. The strongest use case is controlled agentic coding where sovereignty or predictable infrastructure matters. Compare its actual patch quality and total operating cost against both open and hosted alternatives before standardizing, because the surrounding harness and safety design will determine much of the result.
Visit North Mini Code website ↗Similar Coding Tools
View all Coding AI tools →
Devin Desktop
Devin Desktop - Cognition's Windsurf rebrand for managing all your coding agents in a single platform
Kimi K2.7 Code
Kimi-K2.7-Code - Moonshot's open-source coding model, 30% leaner on reasoning
Antigravity 2.0
Antigravity 2.0 - Google's standalone desktop app for orchestrating parallel agents
GLM 5.2
GLM 5.2 - Z AI's new flagship coding model with usable 1M context
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.