Signed prereleases b11192 through b11201 update the HTTP library, add Hexagon binary inspection, add an OpenCL A8 Q8_0 kernel, introduce tiled CPU matrix multiplication for K-quants, repair Jinja compilation and sameas handling, and revert an automatic unified-KV maximum-context adjustment.
Why it matters: The sequence mixes a potentially large CPU throughput change with template correctness and an explicit rollback of context-fitting behavior. Each can affect output, memory or serving reliability on exact workloads.
Signed prereleases b11047-b11056 fix CUDA sort corruption, Metal and JSON-schema handling and several Hexagon operations. Build b11057 then adds a dedicated Ling 3.0 parser so a tool call that arrives before a closing think block is emitted as a tool call instead of being swallowed as reasoning content and terminating an agent loop.
Why it matters: The new parser addresses an agent-loop correctness failure rather than a cosmetic difference. It also extends an already broad prerelease train, so upgrading for Ling 3.0 should remain scoped to the affected chat template and the exact accelerator stack in…
Later signed llama.cpp prereleases on 16 September added a documented fix for an unauthenticated, network-reachable use-after-free in RPC graph caching that could leak libc addresses and enable remote code execution. The sequence through b11007 also repaired int16 transpose corruption on SpaceMiT hardware and retained the earlier OpenCL, ROCm and Hexagon backend changes.
Why it matters: An exposed llama.cpp RPC server could turn a performance cache into a host-compromise boundary, while the SpaceMiT defect could silently corrupt most values in a 4-by-4 transpose. These are security and correctness reasons to replace the earlier 16 September…
Signed llama.cpp prereleases b10919 through b10936 add multi-device Qualcomm Hexagon support, WebGPU and OpenCL correctness fixes, server download-state repairs, an internal JSON-schema representation, structured JSON logging and better complex-type parsing for Qwen3-Coder tool calls. Builds b10932-b10936 also repair cached clang precompiled headers and Jinja integer-property parsing.
Why it matters: The sequence now spans both accelerator correctness and the schema-to-grammar path used for structured tool calls. Local-model operators need to qualify output, tool arguments and observability together instead of treating a successful model load as…
llama.cpp's 8 September prerelease sequence now runs through b10867. Later tags recover a reported Vulkan prompt-processing regression, fix Metal quantized-model widths, protect recurrent checkpoint eviction and disable automatic lazy tensor loading on integrated GPUs after a regression.
Why it matters: These late same-day tags change the risk profile of the existing backend release: they add corrective performance and state-management work while reverting an optimization that regressed integrated-GPU behavior.