TensorFold 1.3.0 is out 🚀
GLM-5.3-Flash went from 83.8 to 104.1 tok/s on two RTX PRO 6000 Blackwells. That’s about 24% faster on CUDA at TP=2 with NVFP4 experts, drafts off and the same output bits in that test.
- On Metal, DeepSeek-V4.1’s prompt-attention kernel is about 1.5× faster, and GLM-5.3-Flash follow-ups reach the first token 3–4% sooner.
- On CUDA, Flash Next frees about 3 GiB of page cache after loading the NVFP4 export. On Linux, you can opt in to mapping KV pages as streams need them.
- tensorfold models now show which chat checkpoints fit your machine. On Metal, tensorfold info accounts for context and parallel streams when estimating memory for Qwen3.8-27B-class models.
- Idol’s denoiser runs natively on Metal and on supported NVIDIA Blackwell GPUs.
- The server has opt-in JSON request logs and a journal for replaying requests after a crash. Stored Responses are scoped to the API key that created them.
- Desktop now loads Qwen3.8-27B correctly from the model library, and TensorFold Web accepts OpenAI-compatible base URLs ending in /v1.
Massive thanks to everyone writing code, testing it on their hardware and filing bugs.
@AIFlow_ML @andersjw_ @arminxnova @atasrb @Bizuayeu @codengod @devopsfortunato @Eleqtrizit @unigilby @GumbiiDigital @jayleaton @teleogen @lasse_ai @mvanhorn @niklaslenz_ai @petruspennanen, and on GitHub AbrahamCete, BadAd84, jschmied, mrv777, supernovae, ultradaoto, vdalesandro, terrapinelf and Mirrdhyn. I really appreciate the time you all put into this.
@petruspennanen @plotarmordev @volatilemarkts @MiaAI_lab @jayleaton