Mellum2 Goes Open Source: A Fast Model for AI Workflows

Source: JetBrains AI Blog, June 1, 2026. Authors: Anton Semenkin, Nikita Pavlichenko.

Note: The New Stack article by Paul Sawers could not be fully retrieved (paywall/interstitial). Content below is drawn from the JetBrains blog post (primary source) and web search results.

What Is Mellum2?

Mellum2 is a 12-billion-parameter model trained from scratch by JetBrains, released as open source under Apache 2.0. It evolves from the original Mellum (4B dense, code completion only) into a full coding assistant capable of natural language interaction, code generation, tool calling, and multi-step agentic workflows.

Architecture

• Mixture-of-Experts (MoE): 12B total parameters, 2.5B active per token. 64 experts, 8 active per token.
• Grouped-Query Attention: 4 KV heads for memory efficiency.
• Sliding Window Attention: Applied on 3 of every 4 layers.
• Multi-Token Prediction (MTP) head: Serves as both auxiliary pre-training objective and built-in draft model for speculative decoding.
• Training: Muon optimizer, FP8 hybrid precision, ~10.6 trillion tokens, three-phase curriculum.
• Context window: 131,072 tokens (up from 8,192 in original Mellum).
• Not multimodal: Trained exclusively on natural language and code data.

Three Variants

All Apache 2.0: 1. Base — Foundation model 2. Instruct — Direct-answer mode for fast agentic commands 3. Thinking — Produces explicit reasoning traces before answering

Performance

• EvalPlus (thinking variant): 78.4% — ahead of Qwen3.5-9B (71.8%) and Seed-Coder-8B (73.8%)
• Single-request throughput: Matches Qwen2.5-7B (~192 tok/s on one H100)
• Under concurrent load: 21% ahead of Qwen2.5-7B, 79% ahead of Qwen3-8B
• Competitive with open-weight models in 4B–14B range while running at per-token compute of a 2.5B dense model
• Benchmarks: LiveCodeBench, BFCL V4, AIME 2025/2026, GSM-Plus, GPQA Diamond, MMLU Redux, JetBrains Internal Pairwise, MixEval, IFEval

Use Cases

1. Routing and orchestration — Analyzing prompts to select the right model or tool
2. Low-latency RAG pipelines — Context retrieval, summarization, response generation
3. Fast sub-agents — Breaking agent pipelines into discrete steps (context gathering, planning, validation)
4. Private/local deployment — Self-hosting for data sovereignty, air-gapped environments

The "Focal Model" Philosophy

"Frontier models will continue to push the limits, but practical AI products also require focal models: fast, specialized components that handle high-frequency tasks efficiently."

The thesis: "the future belongs to coordinated systems, not single models." Performance bottlenecks have shifted from raw capability to latency, throughput, and cost at scale. Many production steps are repetitive, latency-sensitive, and high-frequency — benefiting from a fast, routable model rather than a frontier general-purpose LLM.

Why Open Source

"Open source is how better tools get made."

Available on Hugging Face at the JetBrains collection page.

Technical Report

Companion paper on arXiv: https://arxiv.org/abs/2605.31268