Qwen3.8-27B-Texelator-AWBC4-BF16Cal

Qwen3.8-27B Texelator AW-BC4 (BF16 calibrated, SM120)

This is the standalone Texelator artifact for Qwen3.8-27B, encoded from the original BF16 checkpoint with activation-aware, hardware-exact BC4 PTQ. It is intended for the Texelator v0.3 runtime on NVIDIA RTX 50-series GPUs with compute capability 12.0 and a matching measured BC4 palette hash.

Runtime: https://github.com/hp0303/Texelator

Use

texelator pull qwen3.8:27b
texelator benchmark qwen3.8:27b
texelator run qwen3.8:27b

The one-time benchmark is mandatory. It selects correctness-checked decode lookahead and BF16 prefill tiles for the actual GPU.

Numerical and hardware scope

• Source and calibration dtype: BF16
• Runtime activation/output dtype: BF16
• Accumulation: FP32
• Weight payload: 8-byte BC4 per 16 weights plus one FP32 scale per output row
• Calibration: WikiText-2 train, 8,192 tokens
• Validated target: NVIDIA RTX 5080, compute capability 12.0
• Palette SHA256: 34670bd1c521c26b14ea88ed6f2d8f16c55720f20915909ac621dcabe0f3ed47

Texelator verifies the local GPU's hardware reconstruction palette before loading this artifact. It will fail rather than run on a mismatched palette or architecture.

Prefill modes

BF16 CUTLASS Texture IteratorB prefill is the production default. Texelator v0.3 also includes an opt-in --prefill-mode fp4 research path. FP4 is faster in the reported 512-token experiment but is not quality-equivalent to BF16 and should not be used when output fidelity is required.

License

The upstream Qwen model license and acceptable-use terms continue to apply. The Texelator runtime is MIT licensed.