MTP3 was checked alongside the baseline on B300 DEP8: AIME25: 0.9500 → 0.9667 GPQA: 0.9268 → 0.9242 OCRBench: 0.896 → 0.895 Two validation rounds completed with zero request errors. These are results from the tested build, not a claim that final-release validation is complete. 4/6

Baidu killed the MTP module everyone copied from DeepSeek ERNIE 5.5 Flash just runs its own last layer a second time to guess the token after next. They call it ECHO. Trained on only 1 extra token and it still drafts several ahead for speculative decoding, w/ higher acceptance than an MTP module trained the same way With a normal MTP module the future token loss mostly trains the bolted on module. ECHO puts that loss on the real model so the backbone itself learns to look ahead, and Baidu says reasoning and code got clearly better. The leftover ECHO layer is like 10 times smaller than the MTP module tencent bolted on Hy3, a model the same size

llama.cpp adds MTP for the Qwen3.6 family This is a significant milestone for the local AI ecosystem. The performance jump with these changes is massive and elevates local inference on commodity hardware further. Special thanks to Aman Gupta for leading this development!

We released experimental MTP Qwen3.6 Unsloth GGUFs! Qwen3.6 27B MTP now runs at 140 tokens/s. Qwen3.6 35B-A3B MTP gets 220 tokens/s generation on a single GPU. Qwen3.6 27B and 35B-A3B have >1.4x speed-up over the original GGUFs without any change in accuracy. Guide + GGUFs + Benchmarks: In terms of average speedup, we see a 1.4x for dense models at draft tokens = 2 and for the MoE around 1.15 to 1.2x. We do not recommend more than 2 draft tokens because the acceptance rate drops precipitously from 83% to 50% with 4 draft tokens, and the forward passes for MTP become less beneficial. Use `--spec-type mtp --spec-draft-n-max 2` Thanks to Aman for

Google dropped MTP versions of Gemma4. Ran them on my DGX Spark. The 31B dense model went from 3.94 → 8.91 tok/s. That's +126%. Full results: [26B A4B] > 25.24 → 31.69 tok/s (+25.6%) > TTFT 755 → 332ms (-56%) [31B] > 3.94 → 8.91 tok/s (+126%) > TTFT 599 → 378ms (-37%) If you're not running MTP, you're leaving free perf on the table.