Repository navigation
Releases: LL4nc33/llama-tq
Release list
llama-tq v2026.10.13
Changes: CHANGELOG.md
Linux (x86-64):
- Ubuntu x64 (CPU)
- Ubuntu x64 (Vulkan)
- Ubuntu x64, CUDA: Turing, RTX 20 / T4 (+ PTX for newer GPUs) · Ampere, RTX 30 / A100 · Ada, RTX 40 / L4 · Blackwell, RTX 50 / RTX PRO 6000 (CUDA 12.9)
Linux (arm64):
Docker (CPU, amd64): ghcr.io/ll4nc33/llama-tq:server (also full / light).
Pick the CUDA archive for your GPU generation; the Turing archive also runs on newer GPUs through PTX (slower first start). Other architectures: build from source.
The CUDA archives need the CUDA 12 runtime libraries libcudart.so.12 and libcublas.so.12 (e.g. the packages cuda-cudart-12-8 and libcublas-12-8 from NVIDIA's repository) and a driver for CUDA 12.8 (570 or newer); the Blackwell archive is built with CUDA 12.9 and needs the 12.9 runtime (cuda-cudart-12-9, libcublas-12-9) and a driver for 12.9 (575 or newer).
llama-tq v2026.10.06
Changes: CHANGELOG.md
ci: build only what is on the public branch, when its code changes
Release and Docker images now build on every push to turboquant that touches
code or the WebUI bundle (docs-only pushes build nothing); a newer push
cancels the build of an older one. Removed the scheduled runs inherited from
upstream: daily Docker rebuild, hourly build cache, daily winget, weekly
cross and MSYS builds (still startable by hand).
Linux (x86-64):
Linux (arm64):
Docker (CPU, amd64): ghcr.io/ll4nc33/llama-tq:server (also full / light).
The CUDA build covers sm_75 (Turing) natively plus PTX for newer GPUs; for other architectures build from source.