llama.cpp Adds Support for Tencent Hunyuan 3 MoE Model

The `llama.cpp` project now supports Tencent's Hunyuan 3 (Hy3) architecture, a Mixture-of-Experts (MoE) decoder stack with specific features like per-head Q/K RMSNorm and a sigmoid router. This integration includes MTP speculative decoding and is ported from charlie12345's fork, adapted for mainline APIs.

So What

Tencent's MoE model is now accessible for local inference on diverse hardware via `llama.cpp`, broadening open-source model deployment options.

View Repo→

Sources (2)

llama.cpp Adds Fused Top-K MoE for SYCL to Optimize Inference

PRIMARYllama.cpp Releases · RSS · 1w ago

View repo→

llama.cpp Adds Support for Tencent Hunyuan 3 MoE Model

llama.cpp Releases · RSS · 1w ago

View repo→

Credibility

Lead sourcellama.cpp Releases

Corroboration1 distinct outlet

VerificationSingle source

Entities

Tencentmentioned Anthropicmentioned Intelmentioned

Unlock the full signal with Pro

Tessara synthesizes 80+ data sources into actionable intelligence for AI infrastructure investors.

Unlock with Pro→Sign in

← SignalsSignal

Notable4.8Model ReleaseModelsSingle sourceactive

Updated 1w ago · 2 sources

llama.cpp Adds Support for Tencent Hunyuan 3 MoE Model

The `llama.cpp` project now supports Tencent's Hunyuan 3 (Hy3) architecture, a Mixture-of-Experts (MoE) decoder stack with specific features like per-head Q/K RMSNorm and a sigmoid router. This integration includes MTP speculative decoding and is ported from charlie12345's fork, adapted for mainline APIs.

So What

Tencent's MoE model is now accessible for local inference on diverse hardware via `llama.cpp`, broadening open-source model deployment options.

View Repo→