Code📅 Released: 2026-06-02

JetBrains Mellum2 (12B MoE)

An open-weight 12B parameter MoE model from JetBrains, optimized for IDE integration, routing, and sub-agent workflows.

Overview

JetBrains Mellum2 is a highly compact, 12B parameter Mixture-of-Experts (MoE) coding model released in early June 2026. Specially engineered to power IDE features, it is trained directly on syntax trees, git logs, and software architectures. It excels at low-latency code completion, localized refactoring, and managing micro-agent routing inside development environments.

Unique Factor

A compact 12B MoE model optimized for developer IDEs, providing sub-second code suggestions and smart AST routing.

Key Capabilities

●

IDE Native Optimization

●

12B MoE Compact Size

●

131K Context Window

●

Sub-agent routing

Benchmarks

MMLU Score

80%

HumanEval (Coding)

88.5%

GPQA Diamond

70%

MATH Benchmark

75%

Top Use Cases

IDE Inline Code Completion

Deliver sub-second inline line completions and docstring generations based on neighboring tabs.

Example: “Complete this React hook using the adjacent API service file as context.”

Localized Unit Test Generation

Inspect single modules or classes and write test suites that cover edge cases.

Example: “Write Jest test cases for this utility file. Cover all error handling states.”

Detailed Features

Compact 12B MoE Architecture: Light resource footprint optimized to run on standard developer workstations or local servers.

IDE Native Optimization: Built with special tokens representing code selections, file tabs, and package namespaces.

131,072 Token Context Window: Large enough to ingest multiple source files and configuration scripts simultaneously.

Sub-agent Task Routing: Smart prompt classification designed to divide large tasks among micro-agents.

✓ Strengths & Pros

• Extremely low latency for interactive IDE use
• Open weights allow for local, private deployment on standard hardware
• Fine-tuned specifically on developer AST patterns

✕ Limitations & Cons

• Lower general knowledge reasoning compared to larger models
• Strict coding focus makes it unsuitable for creative writing or business math

Ideal Usage & Target Audience

Best For

Software developers, IDE tool creators, and teams hosting local private coding models.

Not Recommended For

Users performing customer service automation, writing copy, or analyzing financial ledgers.

API Implementation

javascript

const response = await fetch('http://localhost:11434/api/generate', {
  method: 'POST',
  body: JSON.stringify({
    model: 'mellum2-12b-moe',
    prompt: 'Write a typescript interface for a User object.'
  })
});

Check the official documentation for full SDK details.