b10238

ggml-org/llama.cppb10238Aug 3, 2026by github-actions[bot]same day as b10237

What changed since b10237

medium risk

b10238 adds MTP support for Qwen3-Next, extending the MTP work started for DeepSeek V3.2 in b10237.

New features

  • MTP (multi-token prediction) support for the Qwen3-Next architecture, including new load_mtp flags.
  • GGUF tooling now derives num_mtp directly from the MTP layer count for Qwen3-Next models.

Improvements

  • Refactored opt_num_mtp_layers into the _QwenMtpMixin class and cleaned up related comments.
  • Resolved Python type-checking issues across the gguf-py constants and Qwen3-Next model code.
  • Tightened integration in src/models/qwen3next.cpp following review feedback.

Bug fixes

  • Corrected num_mtp computation so it is read from the actual MTP layers rather than inferred indirectly.

Upgrade notes

This is an additive model-support release with no config or CLI migration required; existing DeepSeek V3.2 MTP setups from b10237 are unaffected. Users running Qwen3-Next MTP checkpoints should re-convert or verify their GGUF metadata so the MTP layer count is picked up correctly.

Analyzed Nov 1, 2026

Upgrading?

Release notes summary

Implemented MTP support for the Qwen3-Next model, including fixes to Python type-checking and logic for computing the number of MTP layers.

Key Highlights

  • Added MTP support for Qwen3-Next
  • Fixed Python type-checking issues
  • Implemented logic to compute num_mtp directly from MTP layers
  • Refactored opt_num_mtp_layers and comments

Release Notes (excerpt)

<details open> model: MTP support for Qwen3-Next (#25589) * mtp for qwen3nex * fix for python type-check * Fix to compute num_mtp from directly mtp layer * define opt_num_mtp_layers in _QwenMtpMixin and fix some comments * Fix for python type check * Update gguf-py/gguf/constants.py Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co> * rebase and add load_mtp flags * Update src/models/qwen3next.cpp Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories