Strata adds MoE caching to llama.cpp, claiming 5-10x faster prefill

On October 3, 2026, a post appeared on the LocalLLaMA subreddit introducing a community project called Strata that implements MoE caching for llama.cpp, the local LLM inference tool. The author points to https://github.com/Niko1221/Strata and claims the addition delivers a 5-10x speedup in prefill and a 3-4x speedup in decode.

In the same post, the author says they had been asking the llama.cpp project to add MoE caching for about a year. According to the post, numerous pull requests for the feature moved slowly, delivering only incremental gains in the range of 1% to 2%, even though dozens of papers on arXiv had already demonstrated the concept. The author says Strata produced its implementation in roughly two weeks.

That is where the matter currently stands. The performance figures, the two-week timeline, and the characterization of the earlier pull requests all come from the author's own account. The source material contains no response from the llama.cpp project or any other party, and no independent replication or benchmark result.