TensorFold 1.3.2 is out ๐
96 merged PRs in this release and 31 issues closed since 1.3.0. A lot of that came from people sending patches, trying builds on their own hardware and coming back with an actual repro when something broke. Thank you.
- DeepSeek-V4.1 now takes images on Metal. Decode was 6โ17% faster at one and two streams on an M5 Ultra, with the same tokens in that test.
- Hy4-flash on Metal is included. dots.mocr now takes images on CUDA, and the OCR endpoint reads PDFs on HIP.
- Muse Spark Open runs on two GB10s, and Qwen3.8 Flash Next runs with MTP and prompt snapshots on Intel Xe2.
- Idolโs octree and Gaussian decoders now run natively on M5 Metal.
- tensorfold chat streams replies in your terminal, tensorfold log shows request timings, and you can pull gated Hugging Face models.
Massive thanks to
@AIFlow_ML @andersjw_ @bikepresentball @Bizuayeu @devopsfortunato @unigilby @GumbiiDigital @javierbits @jayleaton @joshuaswarren @teleogen @niklaslenz_ai @plotarmordev @shantanugoel @Shoestring_Dev @arthurostapenko, and on GitHub BadAd84, BobClawblaw, davidzogg13-boop, ElderOrb, hobbesd, jamesadcock, jschmied, SergiioB, supernovae, trickyfalcon, ultradaoto, yguanrong and ercumentyildirim. I really appreciate the time you all put into this.
@plotarmordev @jayleaton @MiaAI_lab @volatilemarkts @petruspennanen