ggml 1.0.0 Released with FP8 and NVFP4 Matrix Operation Support

Posted Nov 14, 07:02 As of Nov 13, 2026

The ggml 1.0.0 release notes add OCP FP8 support and an NVFP4 gate-up matrix multiplication integration that includes GLU. ReBAR detection and a fix for 64-bit FastRPC handle truncation are also included.

What happened

Based on the ggml-org GitHub release notes (2026-11-13)
Added OCP FP8 support to the MMQ path
Added a fused operation that combines NVFP4 gate-up matrix multiplication with GLU
Added automatic libhtp loading
Added ReBAR detection
Fixed a 64-bit FastRPC handle truncation issue
Changes on the llama.cpp side are not directly confirmed in these notes

Our take

Mac (Apple Silicon): The notes alone do not confirm any noticeable change on Mac.
PC: FP8 and NVFP4 only matter with supporting hardware, and the range of supported GPUs is not stated in the notes.
Glasses: The notes contain nothing related to glasses devices.
Unverified: There are no performance figures, and the detailed llama.cpp v1.0.0 changes have not been checked against the original source.

Sources and related posts

ggml-org ggml v1.0.0 release Primary source

Ollama v0.40.17-rc0, Concurrent MLX Generation and iGPU mmap Loading (github.com)

The Ollama v0.40.17-rc0 prerelease includes concurrent generation for MLX and Vulkan mmap loading for shared-memory iGPUs. The OpenAI-compatible endpoint no longer overrides the model's sampler defaults.

Local AI| Nov 14, 07:02 |1 sources |Our take included

llama.cpp b12322 Pre-release, Including Graph Memory, SYCL and CUDA Fixes (github.com)

The b12322 pre-release was published on November 12 at 18:59 UTC, and consecutive builds from b12308 through b12322 were posted the same day. The included changes are the fixes listed on the release page.

Local AI| Nov 13, 07:02 |1 sources |Our take included

Salesforce SOV-Transcribe-1 Speech Transcription Model Weights Released (huggingface.co)

The Hugging Face model card includes instructions for running the model with Transformers and vLLM, along with DFlash speculative decoding draft settings. Benchmarks and parameter counts were not confirmed in the official page body.