Confirmed · Nov 2, 2026, 03:58 UTC

JetBrains releases Mellum2.1-12B-A2.5B-Thinking-MTP with multi-token prediction head

The MTP head works as a speculative draft, with the model verifying every drafted token, so output is unchanged while decoding runs 1.4 to 1.8x faster. About 90% of drafted tokens are accepted, and the head is supported in vLLM 0.33 or newer, transformers 5.20 or newer, and llama.cpp.

• LabJetBrains
• LabHugging Face
• ModelMellum2.1-12B-A2.5B-Thinking-MTP
• ModelMellum2.1 Thinking
• Topicrelease
• Topicopen-weights
• Topicresearch