JetBrains released Mellum2 on June 2, 2026, a specialized coding model designed to run locally on consumer hardware while handling routing, code completion, and summarization tasks that typically require expensive API calls to frontier models. The 12-billion parameter model uses a Mixture-of-Experts architecture that activates only 2.5 billion parameters per token, roughly five times less compute than a traditional dense 12-billion parameter model.

What Makes Mellum2 Different From Other Local Models?

Mellum2 is not positioned as a replacement for Claude, GPT-5.5, or DeepSeek V4. Instead, JetBrains explicitly markets it as a "focal model," meaning a fast, specialized component designed to work inside larger AI systems rather than as a standalone solution. The model is trained exclusively on natural language and code data, deliberately avoiding multimodal capabilities like image or audio processing to maintain specialization for software engineering workflows.

The efficiency gains are substantial. JetBrains reports inference times "less than half" compared to similar-sized models, translating to faster time-to-first-token for interactive use, higher throughput for batch processing, and lower GPU memory pressure for concurrent requests. The model comes with an 8,192-token context window, sufficient for most code completion and summarization tasks, though shorter than frontier models that offer 128,000 tokens or more.

How Can Developers Deploy Mellum2 Locally?

• vLLM Setup: vLLM is the recommended choice for production deployments, with native support for Mellum2's Mixture-of-Experts architecture. Developers can install vLLM via pip and launch an OpenAI-compatible API endpoint at localhost:8000/v1 with a single command.
• Hugging Face Integration: Mellum2 weights are available through Hugging Face under the Apache 2.0 license, allowing developers to load the model using the Transformers library and run inference directly in Python with automatic device mapping.
• Ollama Compatibility: Early community reports suggest Ollama has compatibility challenges with Mellum2's Mixture-of-Experts architecture, though the earlier Mellum-4b dense model works with Ollama. Users waiting for official Mixture-of-Experts support can use the 4-billion parameter dense variant as an interim solution.

Hardware requirements are modest for a 12-billion parameter model. The minimum specification is 16 gigabytes of VRAM, such as an RTX 4080 or A100 40GB GPU, with 24 gigabytes recommended for comfortable batch processing. The Mixture-of-Experts architecture means the memory footprint is larger than a 2.5-billion parameter dense model since all experts must be loaded into memory, but inference remains faster.

What Real-World Tasks Is Mellum2 Built For?

JetBrains highlights several core use cases where Mellum2 excels. Routing and orchestration tasks, where the model analyzes incoming prompts and decides which tool or model handles them, benefit from Mellum2's millisecond-level decision speed. Retrieval-Augmented Generation (RAG) pipeline acceleration is another strength; the model can summarize retrieved context before passing it to a frontier model, keeping the context window smaller for the expensive model and improving coherence.

Fast sub-agents in agentic workflows represent a third use case. In complex workflows where you need quick classification, extraction, or decision-making before the main agent continues, Mellum2 handles intermediate steps without blocking. Teams with data residency requirements or air-gapped infrastructure can run Mellum2 entirely on-premise with no API calls, token billing, or third-party data exposure. High-throughput code features like IDE completions, inline suggestions, and background analysis also benefit from Mellum2's sub-second latency.

The benchmark results position Mellum2 as competitive with other 12-billion parameter models on code tasks. On LiveCodeBench v6, a code generation evaluation suite, Mellum2 performs comparably to similar-sized open-weights models. On AIME 2025/26 mathematical reasoning, GSMPlus grade-school math, GPQA Diamond graduate-level science, and MMLU-Redux general knowledge benchmarks, the model maintains competitive scores while delivering the efficiency advantage that makes it practical for higher-volume deployments.

How Does Mellum2 Fit Into Hybrid AI Workflows?

One emerging pattern is using Mellum2 for initial triage and routing, then handing off complex work to frontier models like Claude Code. Developers can set up a custom Model Context Protocol (MCP) server that routes to Mellum2 locally, allowing Claude Code to offload fast classification, summarization, or extraction tasks to the local model while keeping complex reasoning on the frontier model. This hybrid approach reduces API costs while maintaining reasoning quality where it matters most.

The broader trend reflects a shift in how developers think about local models. Rather than viewing self-hosted models as complete replacements for cloud-based AI services, developers increasingly see them as specialized components in larger systems. Mellum2's design philosophy embodies this shift; it is optimized for speed and efficiency in specific tasks rather than attempting to match frontier models across all dimensions.

Mellum2 is available immediately under the Apache 2.0 license, making it freely usable for both open-source and commercial projects. The release represents JetBrains' effort to provide developers with a practical, efficient tool for local AI deployment without requiring expensive GPU infrastructure or ongoing API costs.