← sources

ggerganov/llama.cpp v1.0.0: v1.0.0

source: github (github:ggerganov/llama.cpp)published 2026-11-13 10:20ingested 2026-11-14 02:34

extract

<img width="1600" height="900" alt="llama-cpp-v1" src="https://github.com/user-attachments/assets/ed94ed55-ebe2-4aee-bdfc-81bdaa3a3f59" /> Blog post: https://huggingface.co/blog/ggml-org/llama-cpp-v1 ## Overview llama.cpp v1.0.0 removes the deprecated public API and promotes the NextN/speculative API to `llama.h`, adds OCP FP8 KV caches, a host-memory MoE expert cache with a configurable CPU share, and loading of external PLE shards. It adds three text architectures (Kimi-K3.1-Pro, Ripple and xing4_0), drama-3 TTS, and MageVL-Cowork and Kimi-K3.1 vision, hardens `llama-server` (SSRF, CORS, DNS rebinding, MCP tool cancellation) and adds a `/v1/audio/speech` endpoint, ships a redesigned web UI with chat groups and a markdown editor, and updates `ggml` to v1.0.0. ### API changes - llama.h: remove `llama_set_warmup`, 35 deprecated functions, `defrag_thold` and `--defrag-thold` ([#30565](https://github.com/ggml-org/llama.cpp/pull/30565)) - llama.h: promote the NextN API from `llama-ext.h` and rename `llama_set_embd_state_src` to `llama_set_ctx_state_src` ([#30565](https://github.com/ggml-org/llama.cpp/pull/30565)) - llama.h: add `moe_cache_size` and `moe_cache_cpu_share` to `llama_context_params` ([#30966](https://github.com/ggml-org/llama.cpp/pull/30966)) - mtmd.h: add `MTMD_GEN_AUDIO_TYPE_DRAMA3` and `temp` in `mtmd_helper_gen_audio_inp` ([#31613](https://github.com/ggml-org/llama.cpp/pull/31613), [#31655](https://github.com/ggml-org/llama.cpp/pull/31655)) - mtmd-helper.h: add `mtmd_helper_resampler` for streaming raw PCM at the model sample rate ([#31964](https://github.com/ggml-org/llama.cpp/pull/31964)) - server: add the OpenAI-compatible `/v1/audio/speech` endpoint and a `temperature` option for `/tts` ([#31763](https://github.com/ggml-org/llama.cpp/pull/31763), [#31655](https://github.com/ggml-org/llama.cpp/pull/31655)) ### New models - Kimi-K3.1-Pro: K3 with the full-attention MLA layers replaced by DSA ([#31779](https://github.com/ggml-org/llama.cpp/pull/31779)) - Ripple-65B-Preview (`ricursive`): new dense 65B architecture, no new ops ([#31692](https://github.com/ggml-org/llama.cpp/pull/31692)) - xing4_0: mHC, MLA, MoE and MTP ([#31589](https://github.com/ggml-org/llama.cpp/pull/31589)) - drama-3 TTS (fishaudio) ([#31613](https://github.com/ggml-org/llama.cpp/pull/31613)) ### Core changes - FP8 KV caches (`-ctk/-ctv f8_e4m3`) with optional static per-tensor scales ([#28392](https://github.com/ggml-org/llama.cpp/pull/28392)) - MoE cache: CPU share for misses, LRFU/ClassLRU replacement, no CPU wait on cached experts ([#30966](https://github.com/ggml-org/llama.cpp/pull/30966), [#31233](https://github.com/ggml-org/llama.cpp/pull/31233), [#31888](https://github.com/ggml-org/llama.cpp/pull/31888)) - external PLE/engram loading from separate GGUF shards via `--shard`/`--ple` ([#31185](https://github.com/ggml-org/llama.cpp/pull/31185)) - Windows unbuffered model load, mirroring the Linux DirectIO path ([#26014](https://github.com/ggml-org/llama.cpp/pull/26014)) - reuse the backend scheduler and graph results across reserves, and skip unused graph metadata pages ([#32009](https://github.com/ggml-org/llama.cpp/pull/32009), [#32089](https://github.com/ggml-org/llama.cpp/pull/32089), [#32043](https://github.com/ggml-org/llama.cpp/pull/32043)) - mark every sampling chain as a graph output to prevent output aliasing ([#32015](https://github.com/ggml-org/llama.cpp/pull/32015)) - remove the non-fused GDN paths and the ubatch-dependent graph node budget ([#32078](https://github.com/ggml-org/llama.cpp/pull/32078)) - store the MTP conv snapshots of every recurrent model in one `set_rows` ([#31660](https://github.com/ggml-org/llama.cpp/pull/31660)) - per-model `build_rope` with a KV-shifting fix and tests ([#31683](https://github.com/ggml-org/llama.cpp/pull/31683)); the model is a member of `llm_graph_context` ([#31733](https://github.com/ggml-org/llama.cpp/pull/31733)) - fix on-device state copy sizes for quantized KV caches ([#30541](https://github.com/ggml-org/llama.cpp/pull/30541)) - follow the reference precision of `ggml_prec_set_acc` call sites ([#31851](https://github.com/ggml-org/llama.cpp/pull/31851)) - attach NVFP4 FFN scales in Granite and qwen2 ([#31449](https://github.com/ggml-org/llama.cpp/pull/31449), [#31657](https://github.com/ggml-org/llama.cpp/pull/31657)); merge tensors on convert for qwen3.5 ([#32102](https://github.com/ggml-org/llama.cpp/pull/32102)) - apply the iGPU lazy-mode fallback only when mmap is off ([#31957](https://github.com/ggml-org/llama.cpp/pull/31957)) - fit: account for skipped NextN layers when setting `n_gpu_layers` ([#31569](https://github.com/ggml-org/llama.cpp/pull/31569)) - jinja: chained comparisons and argument unpacking ([#32072](https://github.com/ggml-org/llama.cpp/pull/32072)); TTS text normalization via the chat template ([#32053](https://github.com/ggml-org/llama.cpp/pull/32053)) - chat: honor `response_format` in the qwen3-coder/deepseek/ministral3, Muse Spark and gemma4 parsers ([#31956](https://github.com/ggml-org/llama.cpp/pull/31956), [#32017](https://github.com/ggml-org/llama.cpp/pull/32017), [#31915](https://github.com/ggml-org/llama.cpp/pull/31915)) - json-schema: fix `allOf` of plain types ([#31834](https://github.com/ggml-org/llama.cpp/pull/31834)); escape control characters and sanitize invalid UTF-8 in tool arguments ([#31937](https://github.com/ggml-org/llama.cpp/pull/31937), [#31861](https://github.com/ggml-org/llama.cpp/pull/31861)) ### Multi-modality changes - remove `clip.h` and consolidate the model/subcontext infrastructure into `mtmd` ([#31865](https://github.com/ggml-org/llama.cpp/pull/31865)) - share the compute buffer among subcontexts to save memory for vision+audio+gen models ([#31945](https://github.com/ggml-org/llama.cpp/pull/31945)) - compute log-mel in the graph instead of spawning threads per request ([#31808](https://github.com/ggml-org/llama.cpp/pull/31808)) - support Gemma 4 streaming audio input ([#31965](https://github.com/ggml-org/llama.cpp/pull/31965)) - support MageVL-Cowork ([#31819](https://github.com/ggml-org/llama.cpp/pull/31819)) and Kimi-K3.1 vision ([#31980](https://github.com/ggml-org/llama.cpp/pull/31980)) - require the projector and patch embedding tensors and a 2x2 spatial merge at load ([#32129](https://github.com/ggml-org/llama.cpp/pull/32129), [#32130](https://github.com/ggml-org/llama.cpp/pull/32130)) - require an F32 position embedding for HunyuanVL and read the MiniCPM-V width from ne[1] ([#32131](https://github.com/ggml-org/llama.cpp/pull/32131), [#32132](https://github.com/ggml-org/llama.cpp/pull/32132)) ### Server changes - security: add `--allow-unauthenticated-tools` and `--media-allow-private-urls` opt-ins ([#32113](https://github.com/ggml-org/llama.cpp/pull/32113), [#32115](https://github.com/ggml-org/llama.cpp/pull/32115)) - security: fix `--cors-origins`, DNS rebinding, CORS proxy SSRF and child instance access ([#32098](https://github.com/ggml-org/llama.cpp/pull/32098), [#32115](https://github.com/ggml-org/llama.cpp/pull/32115), [#32120](https://github.com/ggml-org/llama.cpp/pull/32120)) - restrict the `HF_TOKEN` to huggingface.co hosts ([#31730](https://github.com/ggml-org/llama.cpp/pull/31730)) - cancel MCP tool calls on client disconnect or timeout and notify the MCP server ([#32025](https://github.com/ggml-org/llama.cpp/pull/32025)) - keep the server awake while a request uses its state with `--sleep-idle-seconds` ([#32133](https://github.com/ggml-org/llama.cpp/pull/32133)) - use reference-counted context checkpoints instead of deep copies ([#30137](https://github.com/ggml-org/llama.cpp/pull/30137)) - visit parent slots before child slots for shared-prefix requests ([#30779](https://github.com/ggml-org/llama.cpp/pull/30779)) - realtime API: send error events instead of crashing, reject non-object session updates, fix the output lang ([#31632](https://github.com/ggml-org/llama.cpp/pull/31632), [#31714](https://github.com/ggml-org/llama.cpp/pull/31714), [#31982](https://github.com/ggml-org/llama.cpp/pull/31982)) - accept the top-level `response_format.schema` form and convert Responses `text.format` JSON schemas ([#28697](https://github.com/ggml-org/llama.cpp/pull/28697), [#26145](https://github.com/ggml-org/llama.cpp/pull/26145)) ### UI changes - sidebar: fold, preview and rearrange sections ([#31936](https://github.com/ggml-org/llama.cpp/pull/31936)) and gather chats into named groups ([#32012](https://github.com/ggml-org/llama.cpp/pull/32012)) - move Models, MCP and Settings to a sidebar footer, with full-screen pages on phones ([#32022](https://github.com/ggml-org/llama.cpp/pull/32022)) - mobile: native layout, safe areas and full-screen dialogs ([#31949](https://github.com/ggml-org/llama.cpp/pull/31949)) - open the virtual keyboard in the chat editor on Android ([#32060](https://github.com/ggml-org/llama.cpp/pull/32060)) - chat input: live-preview markdown editor that expands for long messages ([#31955](https://github.com/ggml-org/llama.cpp/pull/31955), [#31973](https://github.com/ggml-org/llama.cpp/pull/31973)) - open the actions menu on right-click for chats, tabs, messages and models ([#31545](https://github.com/ggml-org/llama.cpp/pull/31545)) - PDF attachments: show text and page thumbnails, stop endless conversion retries ([#31468](https://github.com/ggml-org/llama.cpp/pull/31468)) - Discover: show audio capability from the mmproj ([#31837](https://github.com/ggml-org/llama.cpp/pull/31837)) and keep downloads during search ([#31467](https://github.com/ggml-org/llama.cpp/pull/31467)) - add a `search_chats` tool ([#32051](https://github.com/ggml-org/llama.cpp/pull/32051)) and move the JS sandbox setting out of developer ([#32054](https://github.com/ggml-org/llama.cpp/pull/32054)) - take the logo and favicons from the llama-brand package ([#32039](https://github.com/ggml-org/llama.cpp/pull/32039), [#32112](https://github.com/ggml-org/llama.cpp/pull/32112)) - fix a hidden sidebar intercepting clicks ([#31712](https://github.com/ggml-org/llama.cpp/pull/31712), [#32035](https://github.com/ggml-org/llama.cpp/pull/32035)) ### ggml changes - [v1.0.0](https://github.com/ggml-org/ggml/releases/tag/v1.0.0): ggml 1.0.0 prunes the deprecated public API and the legacy `DIAG_MASK` ops, adds configurable YaRN truncation and FP8 support, and brings broad performance work across backends: few-row quantized matmuls and MoE fusion on CUDA/Metal, a Vulkan prefill FlashAttention path, a device-side MoE scheduler on SYCL, decomposed Adreno FlashAttention in OpenCL, new WebGPU kernels, expanded Hexagon op coverage, and scheduler improvements that avoid uploading unused inputs. ## Assets **Nightly build:** [b12359](https://github.com/ggml-org/llama.cpp/releases/tag/b12359) ## More info - [Releases and versioning of `ggml-org` projects](https://github.com/ggml-org/ggml/discussions/1579) ## Changelog since [v0.9.0](https://github.com/ggml-org/llama.cpp/releases/tag/v0.9.0) b1b770d90 examples : remove explict version from test-cmake (#32191) 4cda6227a llama.cpp : bump version to 1.0.0 (#32190) 685981335 sync : ggml 3bfde28f9 ggml : bump version to 1.0.0 (ggml/1690) 28d37fe2e examples : clean-up old/obsolete stuff (#32184) bd2767d06 ci: fix HIP quality check by ignoring the FP8 vec FA kernel spill (#32183) 82503b42f ci : set windows-cuda ccache size to 1G (#32185) 5676dfbe0 chat : accept </ifm|think> as K2 Horizon closer at medium/low effort (#30335) 0a66ac718 server, fix CI: read a small temp file in the read_file tool test [no release] (#32168) 892b40b91 server: allow private media URLs in the vision tests [no ci] (#32167) e97010147 server: skip LAN address tests when the host cannot reach it [no ci] (#32166) 92d6c8a5c fix(hexagon): make host build work with clang-cl on Windows ARM64 (#32011) 8ca5a8008 llama : prune (#30565) 29dbbad2b ggml : prune (#29801) 2a4daf0e0 hexagon: add Q5_0 and Q5_1 support (#32094) d8c8cc4d9 server : cancel MCP tool calls on client disconnect or timeout (#32025) 3ba1ce0e2 qwen3.5: merge tensors on convert also (#32102) 0fb8f1577 Title: CUDA: few-row MMQ for BS 2-24 (#31533) 827466b9e ci : build ui on non-slim runner (#32138) 8954f1948 server: fix SSRF for CORS proxy and media URLs (#32115) a1a68eb79 server: explicit opt-in tools via non-loopback addr, check cors host port (#32113) e182fdad4 server: disallow direct access child instance from browser (#32120) 09452cded server: keep the server awake while a request uses its state (#32133) cfb9d4b75 mtmd: require an F32 position embedding for HunyuanVL (#32131) 72f4603c7 mtmd: read the MiniCPM-V projection width from ne[1] (#32132) 8c388364e mtmd: require the MLP projector and patch embedding tensors (#32129) 08f49598c mtmd: require a 2x2 spatial merge for the Qwen-VL projectors (#32130) a8575a3ef ggml : route GATED_DELTA_NET to the SpacemiT RVV kernel (#31897) a963f6595 vulkan : Intel FA kernel for prefill. (#29357) ab4a77b9a ggml-cpu: use the 4x4 simd_gemm tile on ARM NEON (#32128) 0d69255c5 CUDA: Optimize prefil via fuse of w_s scale in epilogue MMQ for nvfp4 checkpoints (#26048) 3bce8bd50 ggml-backend : prefetch only inputs computed by earlier splits (#32100) 81f21f4b2 ggml : dispatch SpacemiT Q4_1 32x256 through the HP kernel (#31929) b19f1d9e6 ui: add search_chats tool (#32051) d90cfd4a9 ui: install llama-brand from a GitHub tarball instead of git (#32112) 60d46a5ab server: when --tools-runtime is specified, ignore x-tool-runtime (#32119) 93cf736a8 chat : refactor gemma4 to build loose schema rules and clean up (#31915) 4fffab623 llama : avoid touching unused graph metadata pages (#32043) 09338b85e sycl: make oneDNN FA cache thread-safe and memoize runtime failures (#31138) dcbff2166 sycl : fix AOT for more than 4GB (#31873) f3d7725a7 sycl : use cached device-id in set_main_device (#32013) 7021f7e09 CUDA: add cuda debug buffer pool (#31919) e0949d09b opencl: add bin kernel `kernel_gemm_noshuffle_q4_0_f32_32b_trans_ila_a7_bin` (#30947) 5d4c8eeb2 CUDA: skip empty MoE tiles in MMQ stream-k fixup (#32067) dd6b945be server: use the moved test model in the router CORS test (#32117) e1f7593c3 opencl: add decomposed FA for Adreno (#32046) 7bcbc04fb ui: move js sandbox setting out of "developer" (#32054) dd00306bc docs: update image link in README.md (#32038) 1d8199278 ui: keep downloads in Discover during search and fix their progress with a draft (#31467) bd717e08f ui: take the logo and favicons from the llama-brand package (#32039) 2cbe0ec42 server: fix --cors-origins handling + protect against DNS rebinding (#32098) 664f031f4 ci: move test models to ggml-org-dev (#32096) 4180722ee metal : prevent out-of-bounds reads in SOLVE_TRI kernel (#32103) 055c3c26e vulkan: support batched tensor copy API (#31418) 2656b71a1 vendor : update LibreSSL to 4.4.1 [no ci] (#32106) cb44fdb18 llama : reuse graph result instances across scheduler reserves (#32089) cd2db800d jinja : support chained comparisons and argument unpacking (#32072) f7b886fe6 chat: allow TTS model to provide custom text normalizer via jinja (#32053) 956591f45 ci : remove dead snakes (#32084) ccd360f06 llama : remove non-fused GDN and ubatch-dependent graph nodes (#32078) b326dd675 devops : add missing MKL runtime DLLs to the Windows SYCL release package (#32005) b8abe41a5 vendor: update cpp-httplib to 0.66.0 (#32049) 2e325dcd5 backend : re-instantiate CUDA graphs when the cluster dims of a kernel change (#31284) 1648d6411 test-fp8.cpp: use only the public ggml API (#32083) 432433637 ROCm/HIP: fixed BLAS workspace instead of rocBLAS allocating one per call (#31838) 71a6a114b quantize: fix imatrix GGUF filepath metadata leak (#32000) 7f24aacde cpu: accumulate flash attention V in F32 (#29810) af6b5243d webgpu: add PAD_REFLECT_1D support (#31877) d2b3eb9b6 vulkan: allow dequantized KV optimization with CM2 (#31526) eeb210479 vulkan: allocate the device fault vendor binary buffer (#32008) 03d6c824e tests : use only the public ggml API (#32077) adde94c11 ui: open the virtual keyboard in the chat editor on Android (#32060) 459ab1cdd cuda : use MMVF for BF16 matrices of few rows at small batch sizes (#31959) 77e2db8ca cuda : add sm70 FlashAttention config case for DKQ=256, DV=256, ncols=64 (#27997) 9d7b2d739 cuda : fattn mma head sizes 40 72 (#30285) 9e9619d03 CUDA: Add OCP FP8 support (#28413) f3b407ca1 sycl : fixed sycl CLAMP to actually support strided view (#31784) aa75ff323 chat : fix Muse Spark ignoring response_format json_schema (#32017) b5cb8f6f4 ui: keep a hidden sidebar from intercepting clicks again (#32035) 85822a7ab ui: update the README banner to the new llama.cpp cover and remove the legacy brand assets (#31789) 2b66bfbf7 ui: show audio in Discover for models whose mmproj has an audio encoder (#31837) 582eab718 ci : bump ty to 0.0.100 (#32030) 7c2653acc mtmd: fix string format, rename to mtmd_format (#32040) 7592b61d3 ui: move Models, MCP and Settings to a sidebar footer (#32022) cd4ed902e cuda : remove unused ids and type_src0 parameters (#32018) 202c1a229 llama : Add FP8 KV-cache support (#28392) e24d0b4c3 Merge pull request #32015 from ggml-org/gg/fix-sampling-output-aliasing 3e86ce2c2 llama : reuse backend scheduler when sched_params are compatible (#32009) 494d0f47b ggml : add MUL_MAT src1 precision tests and gating (#31896) 281e7dad7 ui: gather chats into named groups in the sidebar (#32012) 90e3841d9 ui: show settings save button even while focused (#32010) c75fba30c cuda : remove unused quantize_cuda_t typedef (#31884) 40f4f4ac9 graph: mark every sampling chain as an output to prevent aliasing 54cde7106 ui: expand the chat form into a large editor for longer messages (#31973) 7d6fd845e convert : mark Sarvam byte fallback tokens as BYTE (#32006) 4ed88da2e convert : workaround conversion for p-e-w/gpt-oss-20b-heretic-ara-v4 (#31860) ac931b5bc cuda : halve VMM pool reservation size and retry on cuMemAddressReserve failure (#30542) 15759e0ca CUDA: skip graph compat check when graph is being reused (#32004) a591cbe17 mtmd: share compute buffer among subcontext (#31945) 203b5f404 server: fix output lang for realtime api (#31982) 351025ec5 mtmd: add kimi K3.1 vision (#31980) 0d9eacadf ggml : fix FreeBSD Vulkan build by setting _XOPEN_SOURCE=700 (#31701) 17701e48d CUDA: move graph and concurrent stream state out of common.cuh (#32001) ac007054e ui: use a live-preview markdown editor for the chat form input (#31955) 940a30ee3 model : stop renaming the recurrent cache tensors in plamo2 (#31984) 45c8ea837 common/json-schema : fix allOf with plain types (e.g. integer) generating an empty object grammar (#31834) c90aa996b common : escape control characters in tool argument strings (#31937) a2c942dcf cann: fail early when checking if buffer type is supported (#31270) 18c5a95e2 sycl: Add device-side MoE scheduler (#30156) ae4893a09 vulkan: add tiled-transpose concat shader for Qwen models (#30591) 6122b566b ggml-rpc : fail gracefully on unreachable endpoints and failed handshakes (#30415) 3837bf0bc vulkan: Divide GROUP_NORM workgroup to 2D to avoid maxComputeWorkGroupCount limit (#31930) c1a4febe9 vulkan: Fix out of bounds reads/writes on MoltenVK (#31890) 78ffbaa18 vulkan: cap scalar flash attention at 8 rows for large heads on Qualcomm (#30829) 58e38bdf6 llama : apply the iGPU lazy-mode fallback only when mmap is off (#31957) d5bbcd1f3 ci: update oneAPI 2026.2, oneDNN 2026.1 and Level Zero SDK 1.34.0 (#31540) 8766c0273 sycl: un-gate oneDNN fused SDPA for Alchemist at head_dim=64 (#31574) 3b4d6ab3f chat : don't use lazy grammar when response_format is set with tools (#31956) cf6463ace hexagon: fix 64-bit FastRPC handle truncation on close (#30952) 711f9f4c1 server: responses api: support conversion of text.format.json_schema to completions (#26145) 26d52b0f9 server

retained text (truncated)

<img width="1600" height="900" alt="llama-cpp-v1" src="https://github.com/user-attachments/assets/ed94ed55-ebe2-4aee-bdfc-81bdaa3a3f59" />

Blog post: https://huggingface.co/blog/ggml-org/llama-cpp-v1

## Overview

llama.cpp v1.0.0 removes the deprecated public API and promotes the NextN/speculative API to `llama.h`, adds OCP FP8 KV caches, a host-memory MoE expert cache with a configurable CPU share, and loading of external PLE shards. It adds three text architectures (Kimi-K3.1-Pro, Ripple and xing4_0), drama-3 TTS, and MageVL-Cowork and Kimi-K3.1 vision, hardens `llama-server` (SSRF, CORS, DNS rebinding, MCP tool cancellation) and adds a `/v1/audio/speech` endpoint, ships a redesigned web UI with chat groups and a markdown editor, and updates `ggml` to v1.0.0.

### API changes

- llama.h: remove `llama_set_warmup`, 35 deprecated functions, `defrag_thold` and `--defrag-thold` ([#30565](https://github.com/ggml-org/llama.cpp/pull/30565))
- llama.h: promote the NextN API from `llama-ext.h` and rename `llama_set_embd_state_src` to `llama_set_ctx_state_src` ([#30565](https://github.com/ggml-org/llama.cpp/pull/30565))
- llama.h: add `moe_cache_size` and `moe_cache_cpu_share` to `llama_context_params` ([#30966](https://github.com/ggml-org/llama.cpp/pull/30966))
- mtmd.h: add `MTMD_GEN_AUDIO_TYPE_DRAMA3` and `temp` in `mtmd_helper_gen_audio_inp` ([#31613](https://github.com/ggml-org/llama.cpp/pull/31613), [#31655](https://github.com/ggml-org/llama.cpp/pull/31655))
- mtmd-helper.h: add `mtmd_helper_resampler` for streaming raw PCM at the model sample rate ([#31964](https://github.com/ggml-org/llama.cpp/pull/31964))
- server: add the OpenAI-compatible `/v1/audio/speech` endpoint and a `temperature` option for `/tts` ([#31763](https://github.com/ggml-org/llama.cpp/pull/31763), [#31655](https://github.com/ggml-org/llama.cpp/pull/31655))

### New models

- Kimi-K3.1-Pro: K3 with the full-attention MLA layers replaced by DSA ([#31779](https://github.com/ggml-org/llama.cpp/pull/31779))
- Ripple-65B-Preview (`ricursive`): new dense 65B architecture, no new ops ([#31692](https://github.com/ggml-org/llama.cpp/pull/31692))
- xing4_0: mHC, MLA, MoE and MTP ([#31589](https://github.com/ggml-org/llama.cpp/pull/31589))
- drama-3 TTS (fishaudio) ([#31613](https://github.com/ggml-org/llama.cpp/pull/31613))

### Core changes

- FP8 KV caches (`-ctk/-ctv f8_e4m3`) with optional static per-tensor scales ([#28392](https://github.com/ggml-org/llama.cpp/pull/28392))
- MoE cache: CPU share for misses, LRFU/ClassLRU replacement, no CPU wait on cached experts ([#30966](https://github.com/ggml-org/llama.cpp/pull/30966), [#31233](https://github.com/ggml-org/llama.cpp/pull/31233), [#31888](https://github.com/ggml-org/llama.cpp/pull/31888))
- external PLE/engram loading from separate GGUF shards via `--shard`/`--ple` ([#31185](https://github.com/ggml-org/llama.cpp/pull/31185))
- Windows unbuffered model load, mirroring the Linux DirectIO path ([#26014](https://github.com/ggml-org/llama.cpp/pull/26014))
- reuse the backend scheduler and graph results across reserves, and skip unused graph metadata pages ([#32009](https://github.com/ggml-org/llama.cpp/pull/32009), [#32089](https://github.com/ggml-org/llama.cpp/pull/32089), [#32043](https://github.com/ggml-org/llama.cpp/pull/32043))
- mark every sampling chain as a graph output to prevent output aliasing ([#32015](https://github.com/ggml-org/llama.cpp/pull/32015))
- remove the non-fused GDN paths and the ubatch-dependent graph node budget ([#32078](https://github.com/ggml-org/llama.cpp/pull/32078))
- store the MTP conv snapshots of every recurrent model in one `set_rows` ([#31660](https://github.com/ggml-org/llama.cpp/pull/31660))
- per-model `build_rope` with a KV-shifting fix and tests ([#31683](https://github.com/ggml-org/llama.cpp/pull/31683)); the model is a member of `llm_graph_context` ([#31733](https://github.com/ggml-org/llama.cpp/pull/31733))
- fix on-device state copy sizes for quantized KV caches ([#30541](https://github.com/ggml-org/llama.cpp/pull/30541))
- follow the reference precision of `ggml_prec_set_acc` call sites ([#31851](https://github.com/ggml-org/llama.cpp/pull/31851))
- attach NVFP4 FFN scales in Granite and qwen2 ([#31449](https://github.com/ggml-org/llama.cpp/pull/31449), [#31657](https://github.com/ggml-org/llama.cpp/pull/31657)); merge tensors on convert for qwen3.5 ([#32102](https://github.com/ggml-org/llama.cpp/pull/32102))
- apply the iGPU lazy-mode fallback only when mmap is off ([#31957](https://github.com/ggml-org/llama.cpp/pull/31957))
- fit: account for skipped NextN layers when setting `n_gpu_layers` ([#31569](https://github.com/ggml-org/llama.cpp/pull/31569))
- jinja: chained comparisons and argument unpacking ([#32072](https://github.com/ggml-org/llama.cpp/pull/32072)); TTS text normalization via the chat template ([#32053](https://github.com/ggml-org/llama.cpp/pull/32053))
- chat: honor `response_format` in the qwen3-coder/deepseek/ministral3, Muse Spark and gemma4 parsers ([#31956](https://github.com/ggml-org/llama.cpp/pull/31956), [#32017](https://github.com/ggml-org/llama.cpp/pull/32017), [#31915](https://github.com/ggml-org/llama.cpp/pull/31915))
- json-schema: fix `allOf` of plain types ([#31834](https://github.com/ggml-org/llama.cpp/pull/31834)); escape control characters and sanitize invalid UTF-8 in tool arguments ([#31937](https://github.com/ggml-org/llama.cpp/pull/31937), [#31861](https://github.com/ggml-org/llama.cpp/pull/31861))

### Multi-modality changes

- remove `clip.h` and consolidate the model/subcontext infrastructure into `mtmd` ([#31865](https://github.com/ggml-org/llama.cpp/pull/31865))
- share the compute buffer among subcontexts to save memory for vision+audio+gen models ([#31945](https://github.com/ggml-org/llama.cpp/pull/31945))
- compute log-mel in the graph instead of spawning threads per request ([#31808](https://github.com/ggml-org/llama.cpp/pull/31808))
- support Gemma 4 streaming audio input ([#31965](https://github.com/ggml-org/llama.cpp/pull/31965))
- support MageVL-Cowork ([#31819](https://github.com/ggml-org/llama.cpp/pull/31819)) and Kimi-K3.1 vision ([#31980](https://github.com/ggml-org/llama.cpp/pull/31980))
- require the projector and patch embedding tensors and a 2x2 spatial merge at load ([#32129](https://github.com/ggml-org/llama.cpp/pull/32129), [#32130](https://github.com/ggml-org/llama.cpp/pull/32130))
- require an F32 position embedding for HunyuanVL and read the MiniCPM-V width from ne[1] ([#32131](https://github.com/ggml-org/llama.cpp/pull/32131), [#32132](https://github.com/ggml-org/llama.cpp/pull/32132))

### Server changes

- security: add `--allow-unauthenticated-tools` and `--media-allow-private-urls` opt-ins ([#32113](https://github.com/ggml-org/llama.cpp/pull/32113), [#32115](https://github.com/ggml-org/llama.cpp/pull/32115))
- security: fix `--cors-origins`, DNS rebinding, CORS proxy SSRF and child instance access ([#32098](https://github.com/ggml-org/llama.cpp/pull/32098), [#32115](https://github.com/ggml-org/llama.cpp/pull/32115), [#32120](https://github.com/ggml-org/llama.cpp/pull/32120))
- restrict the `HF_TOKEN` to huggingface.co hosts ([#31730](https://github.com/ggml-org/llama.cpp/pull/31730))
- cancel MCP tool calls on client disconnect or timeout and notify the MCP server ([#32025](https://github.com/ggml-org/llama.cpp/pull/32025))
- keep the server awake while a request uses its state with `--sleep-idle-seconds` ([#32133](https://github.com/ggml-org/llama.cpp/pull/32133))
- use reference-counted context checkpoints instead of deep copies ([#30137](https://github.com/ggml-org/llama.cpp/pull/30137))
- visit parent slots before child slots for shared-prefix requests ([#30779](https://github.com/ggml-org/llama.cpp/pull/30779))
- realtime API: send error events instead of crashing, reject non-object session updates, fix the output lang ([#31632](https://github.com/ggml-org/llama.cpp/pull/31632), [#31714](https://github.com/ggml-org/llama.cpp/pull/31714), [#31982](https://github.com/ggml-org/llama.cpp/pull/31982))
- accept the top-level `response_format.schema` form and convert Responses `text.format` JSON schemas ([#28697](https://github.com/ggml-org/llama.cpp/pull/28697), [#26145](https://github.com/ggml-org/llama.cpp/pull/26145))

### UI changes

- sidebar: fold, preview and rearrange sections ([#31936](https://github.com/ggml-org/llama.cpp/pull/31936)) and gather chats into named groups ([#32012](https://github.com/ggml-org/llama.cpp/pull/32012))
- move Models, MCP and Settings to a sidebar footer, with full-screen pages on phones ([#32022](https://github.com/ggml-org/llama.cpp/pull/32022))
- mobile: native layout, safe areas and full-screen dialogs ([#31949](https://github.com/ggml-org/llama.cpp/pull/31949))
- open the virtual keyboard in the chat editor on Android ([#32060](https://github.com/ggml-org/llama.cpp/pull/32060))
- chat input: live-preview markdown editor that expands for long messages ([#31955](https://github.com/ggml-org/llama.cpp/pull/31955), [#31973](https://github.com/ggml-org/llama.cpp/pull/31973))
- open the actions menu on right-click for chats, tabs, messages and models ([#31545](https://github.com/ggml-org/llama.cpp/pull/31545))
- PDF attachments: show text and page thumbnails, stop endless conversion retries ([#31468](https://github.com/ggml-org/llama.cpp/pull/31468))
- Discover: show audio capability from the mmproj ([#31837](https://github.com/ggml-org/llama.cpp/pull/31837)) and keep downloads during search ([#31467](https://github.com/ggml-org/llama.cpp/pull/31467))
- add a `search_chats` tool ([#32051](https://github.com/ggml-org/llama.cpp/pull/32051)) and move the JS sandbox setting out of developer ([#32054](https://github.com/ggml-org/llama.cpp/pull/32054))
- take the logo and favicons from the llama-brand package ([#32039](https://github.com/ggml-org/llama.cpp/pull/32039), [#32112](https://github.com/ggml-org/llama.cpp/pull/32112))
- fix a hidden sidebar intercepting clicks ([#31712](https://github.com/ggml-org/llama.cpp/pull/31712), [#32035](https://github.com/ggml-org/llama.cpp/pull/32035))

### ggml changes

- [v1.0.0](https://github.com/ggml-org/ggml/releases/tag/v1.0.0): ggml 1.0.0 prunes the deprecated public API and the legacy `DIAG_MASK` ops, adds configurable YaRN truncation and FP8 support, and brings broad performance work across backends: few-row quantized matmuls and MoE fusion on CUDA/Metal, a Vulkan prefill FlashAttention path, a device-side MoE scheduler on SYCL, decomposed Adreno FlashAttention in OpenCL, new WebGPU kernels, expanded Hexagon op coverage, and scheduler improvements that avoid uploading unused inputs.

## Assets

**Nightly build:** [b12359](https://github.com/ggml-org/llama.cpp/releases/tag/b12359)

## More info

- [Releases and versioning of `ggml-org` projects](https://github.com/ggml-org/ggml/discussions/1579)

## Changelog since [v0.9.0](https://github.com/ggml-org/llama.cpp/releases/tag/v0.9.0)

b1b770d90 examples : remove explict version from test-cmake (#32191)
4cda6227a llama.cpp : bump version to 1.0.0 (#32190)
685981335 sync : ggml
3bfde28f9 ggml : bump version to 1.0.0 (ggml/1690)
28d37fe2e examples : clean-up old/obsolete stuff (#32184)
bd2767d06 ci: fix HIP quality check by ignoring the FP8 vec FA kernel spill (#32183)
82503b42f ci : set windows-cuda ccache size to 1G (#32185)
5676dfbe0 chat : accept </ifm|think> as K2 Horizon closer at medium/low effort (#30335)
0a66ac718 server, fix CI: read a small temp file in the read_file tool test [no release] (#32168)
892b40b91 server: allow private media URLs in the vision tests [no ci] (#32167)
e97010147 server: skip LAN address tests when the host cannot reach it [no ci] (#32166)
92d6c8a5c fix(hexagon): make host build work with clang-cl on Windows ARM64 (#32011)
8ca5a8008 llama : prune (#30565)
29dbbad2b ggml : prune (#29801)

document fetched 2026-11-14 02:34