# llama.cpp version b12206 adds faster MoE caching

> The open‑source LLM runtime releases binaries for macOS, iOS and several Linux builds, improving MoE cache handling so CPU work isn’t stalled.

Oossa · 2026-11-08 · https://oossa.com/en/llama-cpp-version-b12206-adds-faster-moe-caching

ggml‑org released llama.cpp version b12206. The update fixes the MoE (mixture‑of‑experts) cache so the CPU part no longer waits for cached experts. Pre‑built binaries are now available for Apple Silicon macOS, Intel macOS, iOS, and multiple Ubuntu configurations including x64, arm64, and s390x with Vulkan or zDNN support.

## The facts

- Release version b12206 published on 2026‑11‑08
- Adds MoE‑cache improvement to avoid CPU stalls

## Why it matters

Developers can run LLMs faster on their own machines without the CPU bottleneck introduced by earlier MoE caching.

## Sources & references

1. [ggml-org/llama.cpp b12206](https://github.com/ggml-org/llama.cpp/releases/tag/b12206) – llama.cpp, 2026-11-08

Last updated: 2026-11-08