Models

Qwen3.8-27B-BFP4-Infall-TT
Tenstorrent-native BFP4 (bfloat4_b) quantization of Qwen3.8-27B, solved with GPTQ directly on the BFP4 grid. Perplexity rises 4.0% on WikiText-2 versus 11.0% for Tenstorrent's default rounding. For Blackhole.

Inkling-2-Mini-Infall_TT
8-bit and K-quant GGUF builds of Inkling-2-Mini from Thinking Machines Lab (26.4B total, 3.4B active), quantized with the Infall-TT beta preview. Runs on any machine with llama.cpp.