This release ships native Windows and Linux x64 binaries for the RTX 3090, with the current Qwen3.8-27B artifact and the same explicit SM86 runtime profile on both platforms.
- Adds the upstream swscale destination-alignment fix for JPEG decoding, including its 300x200 regression fixture.
- Renders Anthropic and Claude Code system reminders in place after a conversation begins, preserving prefix reuse across turns.
- Streams tool-call whitespace without repeatedly scanning the buffered prefix.
- Ships Windows ZIP and Linux tarball packages with their platform launch helpers.
- iamwavecut — JPEG safety fix, PR #11.
- nasedkinpv — tool-call parser crash fix, PR #12.
- wmehanna — in-place system turns, PR #13.
The release build and smoke gate uses these components:
- Windows: Visual Studio 2022 and CUDA 12.8
- Linux: Ubuntu 22.04, CUDA Toolkit 12.8, GCC 11, CMake 3.28, Ninja, and vcpkg FFmpeg
Each packaged ninfer binary completes a real RTX 3090 Qwen3.8-27B generation before publication.
- This release is a correctness and distribution update, not a new throughput claim.
- Keep the GPU idle while loading the 27B model; a 24 GB RTX 3090 needs the documented explicit KV settings for higher-concurrency profiles.