b10158

ggml-org/llama.cppb10158Jul 28, 2026by github-actions[bot]same day as b10156

What changed since b10156

medium risk

b10158 adds Eagle3-V3 speculative decoding support for GPT-OSS models on top of b10156's HIP precision fix.

New features

  • Added Eagle3-V3 specification support, enabling speculative decoding for GPT-OSS model architectures.

Improvements

  • The HIP -ffast-math disablement introduced in b10156 remains in effect, preserving improved numerical precision on Radeon builds.

Upgrade notes

This is a small incremental step over b10156; no config or CLI changes are documented, so existing invocations should keep working. Rebuild or pull the new binaries only if you need GPT-OSS Eagle3-V3 support, and keep the ROCm/HIP build if numerical precision on Radeon matters to you.

Analyzed Oct 31, 2026

Upgrading?

Release notes summary

Adds support for the Eagle3-V3 specification used in GPT-OSS models, enabling better compatibility with these architectures.

Key Highlights

  • Eagle3-V3 specification support for GPT-OSS
  • Cross-platform binaries maintained

Release Notes (excerpt)

<details open> spec: add eagle3-v3 support for gpt-oss model (#25794) </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - Ubuntu x64 (CPU) - Ubuntu arm64 (CPU) - Ubuntu s390x (CPU) - Ubuntu x64 (Vulkan) - Ubuntu arm64 (Vulkan) - Ubuntu x64 (ROCm 7.2) - Ubuntu x64 (OpenVINO) - Ubuntu x64 (SYCL FP32) - Ubuntu x64 (SYCL FP16) **Android:**…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories