TensorFold 1.2.10: 57% faster Qwen3.8-27B decode on GB10, plus new Metal models
AIFlow_ML · x · 2026-11-19
TensorFold 1.2.10 ships with 57% faster decoding for Qwen3.8-27B on GB10 (8 streams, shared 32K prefix) and 5–7% faster cold prompts for DeepSeek-V4.1 on M5 Ultra. The release adds background responses, tensorfold launch/rm commands, and Qwen3-Embedding plus Unlimited-OCR-2 on Metal. The engine supports Apple Silicon, CUDA and AMD HIP/Vulkan with published real-hardware benchmarks (e.g. Nemotron 3.5 Lightning at 874 tok/s on RTX PRO 6000 Blackwell).
More from Infra
- Exa Boosts Embedding Throughput ~50% on AWS Trainium, Agents Wrote the Kernels — AWS ML Blog · 2026-11-19
- Trump's China data center curb advances despite Amazon, Google, Meta lobbying — pstAsiatech · 2026-11-19
- Liquid models run locally in Mercedes-Benz cars; Shopify needs 20ms query rewriting — agihouse_org · 2026-11-19
- Liquid AI previews on-device continual learning within a fixed memory budget — agihouse_org · 2026-11-19
- Liquid AI keynote: bringing capable AI to the devices around us — agihouse_org · 2026-11-19
- Top 10% OpenAI researchers reportedly burn $25,000 of tokens a day — ivan_bezdomny · 2026-11-19