b9993

ggml-org/llama.cppb9993Jul 13, 2026by github-actions[bot]same day as b9992

What changed since b9992

low risk

Added Hy3 (Hunyuan 3) model architecture and MTP speculative decoding support.

New features

  • Added Hy3 (Hunyuan 3) model architecture support
  • Added MTP speculative decoding support for Hy3

Upgrade notes

No configuration changes are required. The binary now supports the new Hy3 architecture and MTP speculative decoding. Existing GGUFs from the ROCmFPX fork are compatible due to suffix adjustments.

Analyzed Oct 17, 2026

Upgrading?

Release notes summary

Added support for the Hy3 (Hunyuan 3) model architecture, a MoE decoder stack with specific routing mechanisms. It includes support for MTP (Multi-Token Prediction) speculative decoding for this specific model.

Key Highlights

  • Added Hy3 (hy_v3) architecture support
  • Supports MTP speculative decoding for Hy3
  • Ported implementation from ROCmFPX fork
  • Features per-head Q/K RMSNorm and sigmoid router

Release Notes (excerpt)

<details open> model: add Hy3 (hy_v3) support with MTP speculative decoding (#25395) * model: add Hy3 (hy_v3) architecture support Adds Tencent Hunyuan 3 (HF architecture HYV3ForCausalLM, GGUF arch hy_v3): a MoE decoder stack with per-head Q/K RMSNorm, a sigmoid router with expert selection bias, an always-active ungated shared expert, and leading dense block(s) (first_k_dense_replace). The base implementation is ported from charlie12345's fork…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories