Mellum2 Developer Guide: JetBrains' Open-Source Coding Model

Official Sources#

JetBrains released Mellum2 on June 2, 2026 under the Apache 2.0 license. It is a 12-billion parameter Mixture-of-Experts model that activates only 2.5B parameters per token - roughly 5x less compute per forward pass than a dense 12B model. The result is sub-second inference on modest hardware while maintaining competitive benchmark scores for code generation, reasoning, and routing tasks.

This is not a replacement for Claude, GPT-5.5, or DeepSeek V4. JetBrains explicitly positions Mellum2 as a "focal model" - a fast, specialized component inside larger AI systems. Think: routing decisions, RAG summarization, sub-agents, local code completion, and any task where latency matters more than peak reasoning quality.

Last updated: June 18, 2026

What Mellum2 Is Built For#

Mellum2 is trained on natural language and code data. It deliberately avoids multimodal capabilities - no images, no audio - in favor of specialization for software engineering workflows.

The core use cases JetBrains highlights:

Routing and orchestration. Analyze incoming prompts and decide which model or tool handles them. At 2.5B active parameters, Mellum2 can make routing decisions in milliseconds rather than seconds.

RAG pipeline acceleration. Summarize retrieved context before passing it to a frontier model. Keeping the context window smaller for the expensive model saves tokens and improves coherence.

Fast sub-agents. In agentic workflows where you need a quick classification, extraction, or decision before the main agent continues, Mellum2 handles the intermediate step without blocking.

Private local deployment. Teams with data residency requirements or air-gapped infrastructure can run Mellum2 entirely on-premise. No API calls, no token billing, no third-party data exposure.

High-throughput code features. IDE completions, inline suggestions, and background analysis where sub-second latency is required.

Architecture and Efficiency#

Mellum2 uses a Mixture-of-Experts architecture with 64 total experts, activating 8 per token. The full model has 12B parameters, but inference only touches 2.5B per forward pass - a roughly 5x reduction in compute compared to running all 12B.

JetBrains reports inference times "less than half" compared to similar-sized models. In practice, this means:

• Faster time-to-first-token for interactive use
• Higher throughput for batch processing
• Lower GPU memory pressure for concurrent requests

The 8192-token context window is sufficient for most code completion and summarization tasks, though shorter than frontier models. For tasks requiring longer context, you would still reach for a model with 128K+ context.

Benchmark Performance#

JetBrains shared benchmark results for Mellum2 across several evaluation suites. The model competes with similar-sized open-weights models while being faster to run:

The raw numbers position Mellum2 as competitive with other 12B-class models on code tasks, with the efficiency advantage making it practical for higher-volume deployments. It does not match frontier models like Fable 5, GPT-5.5, or DeepSeek V4-Pro on complex reasoning - but that is not the intended use case.

For SWE-bench and similar agentic evaluations, Mellum2 is better suited as a helper model (routing, summarization, tool selection) than as the primary agent.

Local Deployment Options#

Mellum2 weights are available through Hugging Face under the Apache 2.0 license. You can run it locally in several ways:

vLLM (Recommended for Production)#

vLLM supports Mellum2 natively and is the recommended choice for production deployments. The MoE architecture works well with vLLM's optimized inference engine.

pip install vllm

# Download and serve
python -m vllm.entrypoints.openai.api_server \
  --model JetBrains/Mellum2-12B \
  --tensor-parallel-size 1 \
  --max-model-len 8192

This exposes an OpenAI-compatible API endpoint at http://localhost:8000/v1.

Hugging Face Transformers#

For development and experimentation:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "JetBrains/Mellum2-12B"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",
    torch_dtype="auto"
)

prompt = "def fibonacci(n):"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0]))

Ollama Compatibility#

Early community reports suggest Ollama has compatibility challenges with Mellum2's MoE architecture. The earlier Mellum-4b (dense model) works with Ollama, but the new MoE version may require workarounds. Check the Ollama community for current status.

If you need Ollama specifically, consider using the 4B dense variant while waiting for official MoE support:

ollama pull JetBrains/Mellum-4b-base

Hardware Requirements#

For the full 12B model with 2.5B active parameters:

• Minimum: 16GB VRAM (RTX 4080, A100 40GB)
• Recommended: 24GB VRAM for comfortable batch processing
• Quantization: INT8/INT4 quantization reduces memory further if needed

The MoE architecture means memory footprint is larger than a 2.5B dense model (you still load all experts), but inference is faster.

When to Use Mellum2#

Good fit:

• Local code completion where latency matters more than peak quality
• Routing decisions in multi-model pipelines
• RAG context summarization before calling a frontier model
• Sub-agent tasks in agentic workflows
• Private deployments with no API access
• High-volume classification and extraction tasks
• Development and testing before committing to frontier API costs

Not a fit:

• Primary agent for complex multi-step reasoning
• Tasks requiring 128K+ context windows
• Replacing Fable 5, GPT-5.5, or DeepSeek V4 as your main coding model
• SWE-bench-style issue resolution as the sole agent
• Multimodal tasks (images, diagrams, screenshots)

Integrating With Claude Code#

Mellum2 can serve as a fast local model for tasks that do not need frontier reasoning. One pattern is using Mellum2 for initial triage and routing, then handing off to Claude Code for complex work.

Example setup with a custom MCP server that routes to Mellum2 locally:

{
  "mcpServers": {
    "mellum-router": {
      "command": "python",
      "args": ["-m", "mellum_mcp_server"],
      "env": {
        "MELLUM_MODEL_PATH": "/path/to/mellum2",
        "MELLUM_PORT": "8000"
      }
    }
  }
}

This lets Claude Code offload fast classification, summarization, or extraction tasks to Mellum2 while keeping complex reasoning on the frontier model.

Mellum2 vs Other Local Models#

Mellum2's advantage is the combination of code specialization and MoE efficiency. DeepSeek V4-Flash is stronger on absolute capability but has higher active parameters. Qwen3 and Llama 4.1 are competitive but dense, meaning slower inference at similar parameter counts.

FAQ#

What is Mellum2?#

Mellum2 is a 12-billion parameter Mixture-of-Experts model from JetBrains, released June 2, 2026 under Apache 2.0. It activates only 2.5B parameters per token, making it efficient for high-throughput inference while maintaining competitive benchmark scores for code generation and reasoning tasks.

How do I run Mellum2 locally?#

The recommended approach is vLLM for production or Hugging Face Transformers for development. Download the model from Hugging Face (JetBrains/Mellum2-12B) and serve it with vLLM's OpenAI-compatible API server. Ollama support for the MoE architecture is still being worked on.

Is Mellum2 free to use?#

Yes. Mellum2 is released under the Apache 2.0 license, which permits commercial use, modification, and distribution. There are no API costs - you run inference on your own hardware.

How does Mellum2 compare to DeepSeek V4?#

DeepSeek V4-Pro is stronger on absolute capability and benchmarks like SWE-bench. Mellum2 is faster and more efficient for high-volume, lower-complexity tasks. Many teams use both: Mellum2 for routing, summarization, and fast sub-agents; DeepSeek V4 or a frontier model for complex reasoning.

Can Mellum2 replace Claude Code or Cursor?#

No. Mellum2 is designed as a component model, not a standalone IDE agent. It works best alongside frontier models - handling routing, summarization, and fast tasks while the frontier model handles complex reasoning.

What hardware do I need to run Mellum2?#

Minimum 16GB VRAM (RTX 4080, A100 40GB). The full 12B model loads all experts into memory, but inference only activates 2.5B per token. INT8 quantization can reduce memory requirements if needed.

Does Mellum2 support MCP?#

Mellum2 is a model, not a tool. You can build MCP servers that use Mellum2 for inference, and several community implementations exist. The model itself does not have MCP built in - you provide the tool infrastructure.

What context length does Mellum2 support?#

8192 tokens. This is sufficient for most code completion and summarization tasks but shorter than frontier models with 128K+ context. For long-context tasks, use a model with a larger window.

Continue Reading#

• Deploy From Your Coding Agent: Wire Railway''s MCP Server Into OpenCode
• Nimbalyst: A Visual Workspace That Unifies Codex and Claude Code
• Do AI Coding Agents Need Their Own Version Control?

Sources#

• JetBrains Mellum2 Open Source Announcement - June 2, 2026
• Hugging Face Mellum2 Launch Blog
• MarkTechPost: JetBrains Releases Mellum2
• The New Stack: JetBrains open-sources Mellum2
• Neowin: JetBrains open-sources Mellum 2