AI BriefingIK_LLAMA Adds Qwen MTP Support
·2026.04.29 23:22
Key point
IK_LLAMA has added Qwen3.5 MTP support, and MTP-preserved GGUFs are required.
Details
With ik_llama.cpp PR #1698 merged, Qwen 3.5/3.6 MTP support has been added. Current coverage is limited to dense models.
To run it, you need a GGUF with MTP layers preserved. You can build one directly using the qwen-35-mtp-gguf branch, and Qwen 3.5 27B and Qwen 3.6 27B Q8_0 GGUFs for testing have also been released.
Benchmarks were measured under 2x RTX 3090, average of 3 runs, max tokens 4000 conditions.
- Qwen 3.5 27B layer mode: overall rose from 24.8 tok/s to 32.2 tok/s.
- Qwen 3.6 27B layer mode: overall rose from 23.3 tok/s to 29.6 tok/s.
- In graph mode as well, both models showed higher throughput with MTP applied.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Apple Releases LensVLM-9B for Processing Compressed Text Images
Qwen Releases 'RecreationWorld' for Training AI Agents to Rebuild AppsQwen Intelligence Launches Three Mobile AI Agents and Open Benchmark Suite