TensorFold 1.3.2 ships 96 PRs: DeepSeek-V4.1 Metal decode 6-17% faster, 874 tok/s on Blackwell
AIFlow_ML · x · 2026-11-21
Local inference engine TensorFold released v1.3.2 with 96 merged PRs and 31 closed issues since 1.3.0. Highlights: DeepSeek-V4.1 now takes images on Metal with 6-17% faster decode on M5 Ultra at one and two streams; Hy4-flash on Metal; dots.mocr image input on CUDA; OCR endpoint reads PDFs on HIP; Muse Spark Open runs on two GB10s; Qwen3.8 Flash Next runs MTP and prompt snapshots on Intel Xe2; Idol's octree and Gaussian decoders run natively on M5 Metal.
Recorded benchmarks include Nemotron 3.5 Lightning at 874 tok/s greedy decode on RTX PRO 6000 Blackwell and 199 tok/s aggregate across eight streams on two M5 Ultras. Apache-2.0, OpenAI/Anthropic-compatible APIs, Metal/CUDA/HIP/Vulkan support.
More from Infra
- Nebius says Q3 compute deals closed above $25M per megawatt, up from Q2 — Beth_Kindig · 2026-11-21
- NVIDIA lays out 8 design guidelines for sparse attention, reusing selections for 2.4x faster Top-k — NVIDIAAI · 2026-11-21
- llama.cpp adds llama serve: run Inkling 2 Mini GGUF locally with one command — simonguozirui · 2026-11-21
- Ollama CEO in Fortune: enterprise demand for open models keeps growing — ycombinator · 2026-11-21
- Nvidia plunges 13.17% in a single day to $126.75 — jfischoff · 2026-11-21
- vLLM ships Day-0 support for TML model, cuts GB200 decode latency to 3.84ms — vllm_project · 2026-11-21