llama.cpp v0.4.1 has been released, introducing support for several new model architectures and various functional improvements. Key updates include support for the Maple 20B-A1B ternary MoE architecture, Tencent Hy 4 (hy_v4) preview architecture, and Spark2.5.

New Features and Improvements

  • Model Support: Added architectures for Maple 20B-A1B (CPU), Tencent Hy 4, and Spark2.5.
  • Core Enhancements: Refactored JSON schema handling with an internal common_schema representation and split specialized chat parsers into common/parsers.
  • Logging and CLI: Added structured JSONL logging via --log-jsonl and deprecated --mmap, --mlock, and --direct-io in favor of the --load-mode argument.
  • Build Improvements: Added PCH (Precompiled Headers) and unity build support to reduce build times.
  • UI Updates: Improved chat message rendering performance and added UI asset caching.

Bug Fixes

  • Model Corrections: Fixed parameter counts for Granite models and fixed GDN normalization for Qwen, Kimi, and GLM models.
  • Server Stability: Fixed an LRU hang occurring during multiple requests for the same model and resolved speculative decoding issues following multimodal (image) input.
  • Memory and Performance: Fixed MTP context KV cache allocation for DeepSeek2 and GLM-MoE, and disabled lazy tensor loading by default on iGPUs.

Sources (official release notes):