llama.cpp adds runtime support for Prism Bonsai 2 27B
A pre-release build lets the open-source engine load the Hadamard-based model
The facts
- Release b11556 of llama.cpp adds runtime support for the Prism Bonsai 2 27B model (PR #29600)
- The work includes new Hadamard tensor handling folded into a method, plus converter changes and fixes flagged in review
- The change was built with assistance from Claude Code and co-authored by Georgi Gerganov
- Build artefacts ship for macOS, Linux, Windows, Android and openEuler, covering CPU, CUDA 12/13, Vulkan, ROCm, SYCL and OpenVINO backends
llama.cpp has shipped pre-release build b11556 with runtime support for Prism Bonsai 2 27B. It is a 27-billion-parameter model that uses Hadamard transforms in its weights.
The pull request, numbered 29600, folds the Hadamard tensors into a dedicated method. It also adds the conversion logic needed so the model's checkpoint can be loaded and run through llama.cpp's existing inference path.
The changelog shows the patch went through several rounds of cleanup. These included fixes to loading routines and a pass addressing feedback from an automated review bot. Anthropic's Claude Code is credited as an assistant on multiple commits, with Georgi Gerganov, the project's maintainer, listed as a co-author. No benchmark figures, licence terms or model provenance are given in the release notes.
As with every llama.cpp tag, b11556 ships pre-built binaries across a wide range of targets. These span macOS and iOS, Linux on x64, arm64 and s390x, Android, Windows on x64 and arm64, and openEuler. Backend variants cover CPU, CUDA 12 and 13, Vulkan, ROCm, OpenVINO and SYCL.
Why it matters
Adding a model's architecture to llama.cpp is what lets it run on laptops, phones and commodity GPUs instead of a hosted API. Anyone evaluating Prism Bonsai 2 27B can now self-host it across that same broad hardware matrix. That depends on weights becoming available in a compatible format.
Primary source
llama.cpp releases — github.comWritten from the primary source above by The Neural Net’s automated desk (claude-sonnet-5, theneuralnet/pipeline@1) without human review. How this works. Corrections: check the source first — if we got it wrong, the source link is the record.