Hugging Face Daily Digest Oct 21, 2026

Hugging Face's changelog now carries TorchTPU, a native tpu device for Transformers, Diffusers, and Kernels, and Optimum TPU is no longer required for new work. The same overnight posts cover MPS in PyTorch 2.15, with FP8 on Apple silicon. InternScience posted an official Apache-2.0 Q4_K_M llama.cpp pack of Sage-A1.5. Daily Papers flipped to an October 21 stack led by Focused Omni-Modal GUI Agents.

On the Hub

TorchTPU native tpu device

Hugging Face's official changelog now carries TorchTPU, dated 20 October. TorchTPU is a PyTorch plugin that exposes Google TPU as a native tpu device. Transformers, Diffusers, and Kernels support it.¹

New work can use those libraries directly. Install is pip install torch-tpu. Transformers takes device_map="tpu" and Diffusers takes pipe.to("tpu"). The worked example is Qwen3.5-9B, an existing card.

Hugging Face claims Kernels Hub decode 21.5 times faster on Qwen3-30B-A3B and 12.8 times faster on Gemma 4 26B-A4B at batch 64 versus default grouped_mm on TPU v6e-8. Those figures are the post's. This is a library change, not a new Hub checkpoint.

MPS in PyTorch 2.15

Hugging Face's PyTorch team posted MPS in PyTorch 2.15, dated 20 October. FP8 (float8_e4m3fn) lands on Apple silicon with F.scaled_mm. The authors claim MoE decode 5 to 6.5 times faster than 2.14.²

On an M3 Ultra with torch.compile, they report Qwen3.6-35B-A3B at 9.4 to 61.3 tokens per second, and LFM2.5-8B-A1B at 23.6 to 123.6. Wan2.2 tiled VAE decode peaks at about 8 GiB versus 270 GiB.

Requirements are macOS 15 and Python 3.11. The install recipe is the 2.15.0 release candidate, not the default pip torch. Attention fully on Metal is listed as a 2.16 plan. This is a library change, not a new Hub checkpoint.

InternScience Sage-A1.5 official GGUF

InternScience published an official Apache-2.0 llama.cpp pack of Sage-A1.5. The Q4_K_M card was created on 21 October at 04:44 UTC and last modified at 05:03 UTC. The tree is 22.613 GB, with Q4_K_M at 21.713 GB and mmproj at 0.899 GB.³

Sage-A1.5 is a 35B Mixture-of-Experts agentic model from InternScience and Shanghai AI Laboratory. The BF16 parent is still the 15 October card.⁴ Q8_0 and FP8 siblings landed the same morning. InternScience's README says the Q4_K_M files run in llama.cpp, Ollama, and LM Studio.

Daily Papers flip to Oct 21

Daily Papers flipped to an October 21 list. Focused Omni-Modal GUI Agents (arXiv 2610.26203) leads with 101 upvotes. It is an NTU paper with a project page and no Hub model card.⁵

PAR sits at 79 upvotes, with no project page and no Hub card.⁶ Ropedia W0 sits at 64. Its Hub card dates to 18 October and is CC BY-NC 4.0.⁷ Fleet Experience is off today's Daily Papers list of 50.