TensorFold engine 1.2.2 ships ~40% faster prefill on GB10, Metal and HIP gains
HankYeomans · x · 2026-11-04
TensorFold released engine 1.2.2 with notable speedups: Qwen3.8-27B EXL3 prefill is 38-42% faster on GB10 across 2K-64K prompts, Flash Next decode is 36-39% faster on M5 Max, plus faster HIP decode on Ryzen AI Max, native GLM-OCR on HIP, Qwen3.5-9B on Metal, and NVIDIA's Qwen3.8-Flash-Next NVFP4 checkpoint on CUDA TP=2.
The site reports measured runs like 874 tok/s decoding Nemotron 3.5 Lightning on RTX PRO 6000 Blackwell and 199 tok/s aggregate for eight concurrent GLM-5.3-Flash streams on two M5 Ultras. The Apache-2.0 engine supports speculative decoding, prompt reuse and connected Macs, with OpenAI/Anthropic-compatible APIs.
More from Infra
- LlamaIndex speeds up doc-parsing model OCRia-1 10x to 2800 tok/s on one B200 — llama_index · 2026-11-04
- Coding agents more than double every released speculative decoding system, DFlash2 hits 10.87x — berkeley_ai · 2026-11-04
- Scaling Python Without a Dockerfile: Metaflow, DuckDB and Parquet in Practice — anacondainc · 2026-11-04
- Enterprise AI spend hits $177B, up 5.4x; Anthropic and OpenAI take 86% of API spend — deedydas · 2026-11-04
- Chart: CoreWeave burned $2.80 per $1 of revenue at IPO vs Anthropic's $0.14 — NathanpmYoung · 2026-11-04
- Orbital engineers use AI OS CurieOS to draft two-phase cold plate designs — TimothyDuignan · 2026-11-04