# llama.cpp update fixes HunyuanVL position‑embedding crash

> The b12336 release adds a required F32 position embedding for HunyuanVL, preventing buffer overflows on supported platforms.

Oossa · 2026-11-13 · https://oossa.com/en/llama-cpp-update-fixes-hunyuanvl-position-embedding-crash

The open‑source llama.cpp library released version b12336 on 13 Nov 2026. It now forces the HunyuanVL model’s position embedding to be stored as 32‑bit floats (F32) when loading tensors. Earlier, an 64‑bit tensor (F64) could overflow the buffer and crash the first image request. The change applies to all listed platforms, from macOS and Windows to Linux and Android builds.

## The facts

- Release tag b12336 published 13 Nov 2026
- Requires F32 position embedding for HunyuanVL tensors

## Why it matters

Developers using HunyuanVL with llama.cpp will avoid crashes when generating images.

## Sources & references

1. [ggml-org/llama.cpp b12336](https://github.com/ggml-org/llama.cpp/releases/tag/b12336) – llama.cpp, 2026-11-13

Last updated: 2026-11-13