v0.4.5 Preview
Pre-release
Pre-release
Warning
Rolling preview build from v0.4.5 at 476aa6c4b5c6. Assets and moving Docker tags are replaced by newer successful branch builds. Last updated: 01/09/2026 00:09.
Changelog
- Updated the llama.cpp base from
6fdd0ac89to57291f264(b10689). Notable inherited changes include Qwen3.8 Flash Next (qwen4exp) target support with HyperConnections, QSA sparse attention, PLE n-gram embeddings, GGUF metadata and tensor mappings, conversion, multi-GPU placement, and graph lifecycle integration, together with the upstream server, Web UI, conversion, model, and backend updates through that merge point. Bee's release workflows and fork-specific feature surface remain in place. - Added standalone Qwen3.8 MTP sidecar support. The draft loader now uses the configured draft model path, device, and offload settings; the Qwen4Exp MTP graph consumes target hidden states, applies the output HyperConnection mixer, and exports complete sidecars through the converter. Quantized draft caches retain the activation domain required for useful token acceptance.
- Extended standard quantized caches, KVarN, and F16/BF16 precision tails to Qwen3.8 QSA. Sparse selection now composes with the generic attention path instead of bypassing cache writes, native KVarN routing, or precision-tail handling, while the index cache keeps a validated mirrored slot layout.
- Optimized CUDA KVarN decode and speculative verification. Descriptor setup uses a wider reduction and a division-free single-stream scan while preserving host-assigned stage slots; split decode now supports masked query batches through eight rows, skips fully masked splits, reuses decoded K/V across query-row tiles, and selects between 64- and 128-token geometry using occupancy and workload depth. Packed fragment loads, half-width shared quantization axes, and launch bounds reduce decode overhead, while larger or unsupported shapes retain descriptor-native MMA fallback.
- Fixed KVarN compact-plan reuse across cache metadata transitions and padded the CUDA MMA query tile to the physical eight-row
ldmatrixextent. Compact-plan duplicate filtering now uses a capacity bitmap without changing logical cell ordering, and new route, metadata, exact-tail, multi-query, and source-invariant regressions cover the optimized paths. - Hardened Qwen3.8 prompt checkpoints and repeated-request restore. Hybrid attention/index state restores transactionally across ordinary and KVarN caches, precision-tail restores rebuild mirrored slot information, KVarN delegates PLE token-history lookup to its initialized base cache, and PLE recurrent rows restore into the transactionally selected destination.
macOS:
Linux:
- Ubuntu x64 CPU
- Ubuntu arm64 CPU
- Ubuntu x64 CUDA 12.4
- Ubuntu x64 CUDA 13.3
- Ubuntu x64 Vulkan
- Ubuntu x64 ROCm 7.2
- Ubuntu x64 SYCL
Windows:
- Windows x64 CPU
- Windows x64 Vulkan
- Windows x64 SYCL
- Windows x64 CUDA 12.4 - DLLs
- Windows x64 CUDA 13.3 - DLLs
- Windows x64 HIP
Docker:
- CPU:
docker pull ghcr.io/anbeeld/beellama.cpp:server-cpu-preview-v0.4.5 - CUDA:
docker pull ghcr.io/anbeeld/beellama.cpp:server-cuda-preview-v0.4.5 - CUDA 12:
docker pull ghcr.io/anbeeld/beellama.cpp:server-cuda12-preview-v0.4.5 - CUDA 13.3:
docker pull ghcr.io/anbeeld/beellama.cpp:server-cuda13-preview-v0.4.5 - ROCm:
docker pull ghcr.io/anbeeld/beellama.cpp:server-rocm-preview-v0.4.5 - Vulkan:
docker pull ghcr.io/anbeeld/beellama.cpp:server-vulkan-preview-v0.4.5 - SYCL:
docker pull ghcr.io/anbeeld/beellama.cpp:server-sycl-preview-v0.4.5