News · 2026-11-13

llama.cpp reaches version 1.0.0: FP8 KV cache, a host-memory expert cache and server hardening

llama.cpp, the open-source engine behind most local model runners, tagged version 1.0.0 on November 13, 2026. A three-file pull request by Georgi Gerganov bumped the version number, and the release page lists a long changelog since v0.9.0, including FP8 KV-cache support, a host-memory cache for mixture-of-experts weights and a batch of server security fixes. The release body itself says only "More info at 3PM CET."

Key facts

Timeline: pull request #32190 ("llama.cpp : bump version to 1.0.0") was opened at 08:51 UTC and merged at 09:06 UTC on November 13; the llama.cpp v1.0.0 release was published at 10:20 UTC, and the ggml v1.0.0 release went out at 08:34 UTC.
PR size: 3 files, +4/-4.
Not stated: any maintainer statement of what 1.0 promises about stability.

What the changelog supports

The commit list, not a prose summary, is the source here. Items include:

Memory: FP8 KV-cache support (#28392) and OCP FP8 in CUDA (#28413); see our lesson on the KV cache for why that matters.
Mixture-of-experts: a host-memory expert cache with LRFU and ClassLRU replacement (#31233) and CPU fallback for missed experts (#30966), plus external PLE file loading (#31185).
Models: Kimi-K3.1-Pro text (#31779), Ripple (#31692), xing4_0 (#31589), and vision for Kimi K3.1 (#31980) and MageVL-Cowork (#31819).
Speech: drama-3 text-to-speech (#31613) and an OpenAI-compatible /v1/audio/speech endpoint in the server (#31763).
Hardware: a Vulkan Intel prefill FlashAttention kernel (#29357) and Metal MoE routing fusion (#31615).
Server hardening: an SSRF fix for the CORS proxy and media URLs (#32115), --cors-origins handling with DNS-rebinding protection (#32098), cancellation of MCP tool calls on client disconnect or timeout (#32025), and restricting the Hugging Face token to huggingface.co hosts (#31730).

The version-number footnote

The pull request drew one comment, from contributor danbev, noting that examples/test-cmake still asks for find_package(llama 0.1.0). The fix, #32191, removes the explicit version instead of bumping it. This small wrinkle reflects how the project versions itself: llama.cpp had gone through ten thousand build numbers before it tagged its first version (see our earlier note on v0.1.0), so 1.0.0 is a labeling milestone as much as a feature release.

What not to read into it

Some early commentary framed 1.0 as a cleanup removing deprecated API. The release body does not say that, and no changelog item verified here supports it; "prune" commits (#30565, #29801) exist but were not read. Specific claims about counts of deprecated functions and new flag names circulated without support in the release text and were left out. Community reception was not read either.

Why it matters to local-model users

Two items affect what people can run. The expert cache is the same idea behind projects that keep hot experts on the GPU and stream the rest, a pattern our piece on a patch that learns which experts to keep in VRAM described. FP8 KV cache can roughly halve cache memory compared with 16-bit, at some accuracy cost, which is a standard trade-off our lesson on quantization covers. The server fixes matter because llama-server is often exposed on a local network: SSRF and DNS rebinding are the classic ways a web page or a crafted media URL reaches into a machine it should not.

Bottom line

A version number is a promise people read into, and this release does not yet say what the promise is. The changelog is the concrete content: FP8, expert caching, new models, a speech endpoint and a security-minded sweep of the server.

Primary source, verified: read the paper →

Key questions

What changed in llama.cpp 1.0.0?

The release is mostly a version bump plus a long changelog since v0.9.0. Notable items include FP8 KV-cache support, a host-memory cache for mixture-of-experts weights, vision for Kimi K3.1, a text-to-speech endpoint and server security fixes.

Does 1.0 mean the API is now stable?

The release does not say so. No maintainer statement about what 1.0 promises was found, and the release body only says 'More info at 3PM CET' before listing commits.

Why does the expert cache matter?

A mixture-of-experts model activates only a few experts per token, so keeping the frequently used ones in host memory and falling back to the CPU for misses lets larger models run on smaller GPUs.

Cite this

APA

Ground Truth. (2026, November 13). llama.cpp reaches version 1.0.0: FP8 KV cache, a host-memory expert cache and server hardening. Ground Truth. https://groundtruth.day/news/llama-cpp-reaches-version-1-0-0-with-fp8-kv-cache-and-a-host-memory-expert-cache.html

BibTeX

@misc{groundtruth:llama-cpp-reaches-version-1-0-0-with-fp8-kv-cache-and-a-host-memory-expert-cache,
  title  = {llama.cpp reaches version 1.0.0: FP8 KV cache, a host-memory expert cache and server hardening},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {nov},
  url    = {https://groundtruth.day/news/llama-cpp-reaches-version-1-0-0-with-fp8-kv-cache-and-a-host-memory-expert-cache.html}
}

Topics: llama-cpp · local-inference · open-source · release · moe · quantization

Comments are replies to this story on Bluesky — reply with any Bluesky account to join in.