Maple Preview Windows b2

Maple Preview / DeepGrove llama.cpp Windows b2

This release keeps the tested Windows build used for Maple Preview together with two unmodified upstream llama.cpp Windows baseline builds for easy comparison and fallback.

Assets

• llama-b2-bin-win-vulkan-common_cpus-x64.zip

  • Custom DeepGrove Maple-capable llama.cpp Windows x64 build.
  • Vulkan enabled, including the TQ2_0 Vulkan path used by Maple Preview.
  • Jan-compatible archive naming.
  • Includes the OpenSSL runtime DLLs required by llama-server.exe (libssl-3-x64.dll and libcrypto-3-x64.dll).
  • This is the exact packaged b2 build that was tested before publishing this release.

• llama-b10621-bin-win-cpu-x64.zip

  • Official, unmodified upstream ggml-org/llama.cpp Windows CPU x64 build from tag b10621.

• llama-b10621-bin-win-vulkan-x64.zip

  • Official, unmodified upstream ggml-org/llama.cpp Windows Vulkan x64 build from tag b10621.

• SHA256SUMS.txt

  • SHA-256 checksums for all ZIP assets.

Jan

In Jan, install llama-b2-bin-win-vulkan-common_cpus-x64.zip from:

Settings -> Llama.cpp -> Install Backend from File

Then select release version b2 and backend win-vulkan-common_cpus-x64.

Maple Preview note

Maple Preview loaded correctly with this build. In one tested Intel Arc A750 8 GB setup, GPU offload through the experimental Vulkan/TQ2_0 path produced incorrect generations, while GPU Layers = 0 (CPU-only) produced correct output and measured about 91 tok/s on an Intel Core i7-14700. This is a single hardware/runtime observation, not a universal benchmark.

For Intel Arc users, CPU-only is therefore the safer starting point for Maple Preview with this b2 build. Vulkan support should be treated as experimental until validated on the target GPU/driver.

Provenance

This repository is a fork of deepgrove-ai/llama.cpp. The two b10621 baseline ZIPs are redistributed unchanged from the official ggml-org/llama.cpp release for convenience and comparison.