JetBrains/Mellum2.1-12B-A2.5B-Thinking-MTP · Hugging Face
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.
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
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.
JetBrains/Mellum2.1-12B-A2.5B-Thinking-MTP · Hugging Face
huggingface.co · 2 November 2026
Loading the full article…
This text was published by huggingface.co. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
1source- Reddit discussionreddit.com
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- ChatGPT leads AI image generation in blind tests, Gemini best value, Grok fastest · 1 src
- Nace AI releases Drex Max v1, a 27B decision model that runs locally on Apple Silicon · 1 src
- AI News Summary 11/1 | Claude's Ultra 5-hour limit removed, 9 AI infrastructure companies breached, and more · 1 src
- Claude removes Opus 4.8 fast mode, users should migrate to Opus 5.5 · 1 src
- Anthropic renames $200 Claude plan to Ultra and removes 5-hour limit · 1 src
Comments
via GitHub Discussions