TensorFold engine 1.2.2 is out. ๐
Some of what's landed:
- Qwen3.8-27B EXL3: 38% to 42% faster prefill on GB10 across 2K to 64K prompts.
- Flash Next on M5 Max: 36% to 39% faster decode with SSD-backed embeddings.
- Faster HIP decode on Ryzen AI Max.
- NVIDIA's Qwen3.8-Flash-Next NVFP4 checkpoint on CUDA at TP=2.
- Native GLM-OCR on HIP and Qwen3.5-9B on Metal.
- Faster CUDA loading, less duplicate memory after loading on CUDA/Metal, and fixes for malformed checkpoint headers.
- Signed archives are available for macOS arm64, Linux x86_64 and Linux aarch64.
Thank you to all 27 contributors and upstream authors credited in this release, and everyone testing, reporting bugs and lending hardware.
@bhcc2025 @Bizuayeu @BlivionIaG @codengod @diwakergupta @unigilby @grantoverton @jayleaton @jishvenugopal @JRB_AI @teleogen @dennis_rye @2Halves1brain @petruspennanen
GitHub credits: AbrahamCete, Arminova, BadAd84, BobClawblaw, fkiene, garvrao80, jschmied, kieranscotney, optimisme, paragontasx, pauldodd123, ppardus, russellmm.
@petruspennanen @MiaAI_lab @volatilemarkts @plotarmordev @tensorfold_dev