The latest Ollama update fixes a frustrating bug where model downloads would stall before sending any data. It also resolves an issue with GLM tool calls being silently dropped at the end of generation, ensuring that tool interactions are reliable.
Integration improvements include the restoration of Claude Code Channels and a fix for Anthropic thinking streams. The Hermes Desktop client now correctly honors the --force-build flag. Furthermore, Laguna 2.1 models are now fully supported for chat, thinking, and tool calling across all platforms, including a specific fix for Metal inference on Apple devices.
On the GPU front, CUDA support has been extended to Windows ARM64, and compatibility with NVIDIA B200 GPUs is introduced through CUDA 12. Linux systems with CUDA or ROCm iGPUs will see lower memory usage. Finally, both the MLX and llama.cpp engines have been refreshed to their latest versions, bringing general performance and stability improvements.
