JetBrains’ Mellum2.1 Thinking targets local coding agents with 2.5B active parameters
The Apache 2.0 MoE model pairs a 131K context window with official GGUF builds and strong self-reported coding gains, though several Qwen3.5 scores remain higher.
2 min read
JetBrains has released Mellum2.1 Thinking, an Apache 2.0 reasoning model aimed at coding, tool use and repository-scale agent tasks. The mixture-of-experts model has 12 billion total parameters but activates 2.5 billion per token, using 64 experts with eight selected at a time. Its published context window is 131,072 tokens.
The release is available as BF16 weights on Hugging Face, with vLLM serving instructions and an OpenAI-compatible chat example. JetBrains has also published official GGUF builds ranging from a 24.3 GB BF16 file to a 7.0 GB MXFP4 MoE quantization for llama.cpp, Ollama and other local runtimes. Hugging Face does not currently list a hosted inference provider for the main model.
JetBrains reports large gains over Mellum2 under one consistent internal evaluation pipeline. Mellum2.1 scores 82.0 on LiveCodeBench v6 versus 69.4 for Mellum2, while SWE-bench Verified rises from 2.0 to 47.0 and Terminal-Bench 2.1 from 0.6 to 17.4. It also reaches 62.3 on BFCL v4. The same table shows limits: Qwen3.5 9B remains ahead on SWE-bench Verified at 50.0, Terminal-Bench at 21.7, AIME 2025/2026 at 86.7 and GPQA Diamond at 77.8.
All reported numbers are self-reported by JetBrains. Non-agentic tests used greedy decoding, while agentic tests used Pi v0.73.1 with shell and file tools, a 114K-token context and up to 16K generated tokens per turn. That controls the harness across models, but it does not substitute for independent testing, and agent scores can move with scaffolding and sampling choices.
The practical appeal is the efficiency profile: a model with 2.5B active parameters, long context and permissive weights that can run locally in quantized form.
Sources: Mellum2.1 Thinking model card and evaluation · Official GGUF builds
Conversation
0 approved commentsNo comments yet. Start a thoughtful conversation.
Join the conversation.
Sign in with ChatGPT to leave a comment or rate this article.