llama.cpp has released an update to support AMD's RDNA3.5 GPU architecture, specifically targeting the gfx1151 chip found in devices like the Radeon 8060S. The change modifies the MoE (Mixture of Experts) kernel heuristic in ggml-cuda/mmq.cu to recognize and optimize for this newer hardware generation. Testing on an AMD Ryzen AI MAX+ 388 system shows a 16.2% performance increase for Q4_K quantized MoE models compared to previous builds. This update ensures that users with the latest AMD GPUs can achieve optimal inference speeds without manual configuration.
Read originalThis release quietly extends llama.cpp's hardware support to the latest driver stacks, adding official binaries for ROCm 10.0 and CUDA 13 across Linux and Windows. For AMD users, this means native compatibility with newer GPU architectures without manual compilation tweaks, while NVIDIA users gain access to the latest CUDA runtime optimizations. The inclusion of WebGPU in CI signals ongoing work toward browser-based inference, though it remains a background effort for now. There are no new model formats or quantization methods here, just broader infrastructure coverage that keeps llama.cpp relevant as hardware evolves.
A critical precision bug in llama.cpp’s Apple Silicon backend has been patched, resolving total inference failures on models with high-activation ranges like Mistral Small 4. The issue stemmed from f16 saturation during matrix multiplication, which turned entire output tensors into NaN values for inputs exceeding ~32 tokens. By implementing an exact, power-of-two rescaling mechanism in the Metal kernel, the fix restores correctness without significant performance penalties. This ensures local inference on M-series chips remains viable for complex MoE architectures that previously crashed.
This release stabilizes the core agent loop by fixing critical bugs in prompt caching and subagent resume logic that previously broke context reuse. Plugin management gets a significant upgrade with dynamic folder scanning, allowing developers to hot-load tools without restarting the session. The update also hardens security around symlink traversal and refines telemetry routing for enterprise gateways. While not feature-heavy, these fixes make Claude Code more reliable for complex, multi-step coding workflows.
Anthropic quietly patched a regression in Claude Code that broke LLM-gateway and proxy setups. The previous version incorrectly forced Cloud-gateway sign-in when the CLAUDE_CODE_USE_GATEWAY variable was set, even with API keys or custom auth headers configured. This update restores the expected behavior where the variable is ignored unless specific Anthropic credentials are present. Developers relying on custom authentication flows can now resume their workflows without configuration changes.