# JetBrains/Mellum2.1-12B-A2.5B-Thinking-MTP · Hugging Face

Digest AI · Generative AI & Models · published 2026-11-02T03:58:06Z

Canonical: https://digestai.news/story/jetbrains-mellum2-1-12b-a2-5b-thinking-mtp-hugging-face

## Summary

The 12 B mixture‑of‑experts model keeps the same architecture as the previous version—28 layers, 2.5 B active parameters, 64 experts (8 active per token), 131,072 token context length, and a 1,024‑token sliding window—but adds a speculative decoding draft layer trained via reinforcement learning at scale.

The release is open‑source under the Apache 2.0 license and requires vLLM 0.33+ and transformers 5.20+ to use the head. The model is positioned for complex agentic tasks such as repository navigation, command execution, and tool use, as well as hard non‑agentic coding, math, and reasoning problems.

## Key points

- 12 B MoE model with 2.5 B active parameters adds a multi‑token prediction head for speculative decoding
- Self‑reported benchmarks show strong coding and reasoning scores, e.g., 91.500 % HumanEval pass@1

## Why it matters

An open‑weight 12 B model that decodes up to twice as fast can lower compute costs for developers building autonomous agents and coding assistants.

## Sources

1. [JetBrains/Mellum2.1-12B-A2.5B-Thinking-MTP · Hugging Face](https://huggingface.co/JetBrains/Mellum2.1-12B-A2.5B-Thinking-MTP) (huggingface.co, 2026-11-02, primary source)

## Cite

Digest AI, "JetBrains/Mellum2.1-12B-A2.5B-Thinking-MTP · Hugging Face", 2 November 2026, https://digestai.news/story/jetbrains-mellum2-1-12b-a2-5b-thinking-mtp-hugging-face

---

Written by Digest AI's editorial model from the linked sources; the sources are the record. Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse
JSON: https://digestai.news/story/jetbrains-mellum2-1-12b-a2-5b-thinking-mtp-hugging-face.json