llama.cpp b9990 adds Minimax2 Eagle3 speculative decoding support
New version enables faster local LLM inference with speculative draft models.
The latest release of llama.cpp, tagged b9990, introduces support for Minimax2 Eagle3—a speculative decoding architecture designed to accelerate LLM inference. Speculative decoding works by using a smaller draft model to generate candidate tokens in bulk, which are then verified by the main model. This approach can significantly improve throughput for local AI models, especially on consumer hardware.
Alongside the new feature, the release includes a critical fix for a nullptr crash in the minimax2 EAGLE3 code path. The llama.cpp team also added a minor newline improvement. Builds are available across a wide range of platforms: macOS (arm64, Intel, iOS XCFramework), Linux (x64, arm64, s390x, Vulkan, ROCm, OpenVINO, SYCL), Android (arm64), and Windows (CPU, CUDA 12/13, Vulkan, OpenVINO, HIP). This release continues llama.cpp's mission to make cutting-edge LLM techniques accessible for local execution.
- Adds Minimax2 Eagle3 speculative decoding, enabling faster inference with a draft model.
- Fixes a nullptr crash in the minimax2 EAGLE3 code path.
- Builds available for macOS, Linux, Windows, Android, iOS, and multiple GPU backends (CUDA, Vulkan, ROCm, etc.).
Why It Matters
Speculative decoding in llama.cpp makes local LLMs faster and more practical for real-time applications on consumer hardware.