skulldc

The llama.cpp v1.0.0 release hardens the server and adds an FP8 KV cache

ggml-org tagged llama.cpp v1.0.0 with SSRF and DNS rebinding fixes in the server, FP8 KV-cache support, new models and a pruned codebase since v0.9.0.

A short note and a long changelog

ggml-org published llama.cpp v1.0.0 on 2026-11-13. The release notes themselves are brief: they say a new version has been released, add a note that more information follows at 3PM CET, and point to a discussion about releases and versioning of ggml-org projects. The matching nightly build is b12359. The ggml library was bumped to 1.0.0 in the same cycle (ggml/1690), and llama.cpp’s own version bump landed as #32190.

What the notes do carry is the full changelog since v0.9.0. The items below are taken from that list, grouped by who is most likely to care.

Server changes for anyone exposing llama-server

Several commits tighten how the HTTP server handles network access and outbound requests:

  • server: fix SSRF for CORS proxy and media URLs (#32115).
  • server: fix --cors-origins handling + protect against DNS rebinding (#32098).
  • server: explicit opt-in tools via non-loopback addr, check cors host port (#32113).
  • server: disallow direct access child instance from browser (#32120).
  • server: when --tools-runtime is specified, ignore x-tool-runtime (#32119).
  • server : cancel MCP tool calls on client disconnect or timeout (#32025).

Two related fixes sit outside the server. common : restrict HF token to huggingface.co hosts (#31730) limits where a Hugging Face token is sent, and quantize: fix imatrix GGUF filepath metadata leak (#32000) addresses file paths ending up in GGUF metadata. If you run llama-server on anything other than loopback, or let it fetch media and call tools, these are the commits to read before upgrading.

On the API side, the server gains an oai-compat /v1/audio/speech endpoint for text to speech (#31763), and the Responses API path now converts text.format.json_schema to completions (#26145).

Precision, models and the MoE cache

llama : Add FP8 KV-cache support (#28392) and CUDA: Add OCP FP8 support (#28413) both appear in this cycle, alongside a CUDA commit that fuses the scale into the MMQ epilogue for nvfp4 checkpoints (#26048).

New model support includes the Kimi-K3.1-Pro text model (#31779), Kimi K3.1 vision (#31980), Ripple from ricursiveai (#31692) and MageVL-Cowork (#31819). The multimodal library also gained Gemma 4 input audio streams (#31965).

The MoE expert cache received several changes: missed experts now run on the CPU with a tuning script (#30966), a hybrid LRFU/MQ-like cache policy called ClassLRU was added (#31233), and the cache is no longer applied to a speculative draft model (#31497).

Backends, pruning and build notes

Backend work is spread widely. Vulkan adds an Intel flash attention kernel for prefill (#29357) and an RDNA3/4 large-tile split-K optimization (#30280). OpenCL adds decomposed flash attention for Adreno (#32046). Hexagon adds Q5_0 and Q5_1 support (#32094). SYCL gains a device-side MoE scheduler (#30156). ROCm/HIP switches to a fixed BLAS workspace instead of letting rocBLAS allocate one per call (#31838). Metal fixes out-of-bounds reads in the SOLVE_TRI kernel (#32103) and out-of-bounds writes to the flash attention output buffer (#31662).

The release also removes code. Commits titled llama : prune (#30565) and ggml : prune (#29801) appear next to examples : clean-up old/obsolete stuff (#32184) and the removal of non-fused GDN and ubatch-dependent graph nodes (#32078). Anyone building against older examples or internal graph paths should check those diffs.

For packagers, CUDA 13 release builds move to 13.5 (#31788), CI runners move to Ubuntu 26.04 (#31529), and vendored dependencies are updated to LibreSSL 4.4.1, BoringSSL 0.20261105.0 and cpp-httplib 0.66.0.


Source: ggml-org/llama.cpp v1.0.0 release notes