Summary
JetBrains has released Mellum 2, a 12-billion-parameter mixture-of-experts coding model that runs at the compute cost of a 2.5-billion dense model while supporting a 131,072-token context window. Built entirely from scratch—not fine-tuned from another base—and released under the Apache 2.0 license, Mellum 2 adds chain-of-thought reasoning, external tool calling, and multi-step agentic workflows to what was previously a single-purpose IDE completion model. Three of its four attention layers use a sliding window of just 1,024 tokens for efficiency, while the fourth performs full attention to preserve long-range context.
In this step-by-step tutorial, Fahd Mirza deploys Mellum 2 locally on an NVIDIA RTX A6000 with 48GB VRAM using vLLM (consuming roughly 45GB at runtime, largely for KV cache) and connects it to an MCP filesystem server via MCP CLI. With 14 filesystem tools loaded and a Qwen 3 reasoning parser exposing chain-of-thought as a separate API field, the model autonomously lists directories, creates files, reads them back, and chains multiple tool calls without re-prompting. Mirza then connects the model to the Hermes coding agent for additional code generation testing.
The walkthrough is practically oriented—covering the exact vLLM command flags, JSON config structure for the MCP server, NPX-based server startup, and common installation gotchas—making it directly replicable for developers who want a self-hosted open coding agent with real filesystem access and transparent reasoning traces.
📺 Source: Fahd Mirza · Published June 05, 2026
🏷️ Format: Hands On Build
