Hugging Face Daily Digest Oct 20, 2026

Unsloth finished the Dynamic 3.0 GGUF ladder for DeepSeek-V4.1-Flash-1017 overnight. The WIP banner is gone, and the smallest 1-bit set still weighs 233 GB. Tencent posted official llama.cpp files of Hy-Memory-30B-A3B. Hugging Face opened its changelog on TorchNeuron, a native Trainium device for Transformers, Diffusers, and Kernels, with public beta today. Daily Papers flipped to an October 20 stack led by Fleet Experience.

On the Hub

Unsloth DeepSeek-V4.1-Flash-1017 Dynamic ladder

Unsloth posted the rest of the MIT Dynamic 3.0 GGUF ladder for DeepSeek-V4.1-Flash-1017 late on 19 October. The card now holds 89 files and 4060.199 GB, last modified at 23:51 UTC. UD-IQ1_S is the floor at 232.980 GB, with an engram folder of 609.810 GB. Q4_K_XL is 500.708 GB and Q8_K_XL is 508.377 GB. The README now opens on Unsloth's run guide.¹

The repo opened on 17 October. Last night's drop is the rest of the quants, plus a DSpark draft around 8 GB and mmproj files. Unsloth writes that the smallest 1-bit needs 128 GB of RAM or more. It puts 3-bit at 256 GB and full precision at 320 GB. Those figures are Unsloth's.

The official DeepSeek card did not move. Last modified is still 17 October at 05:46 UTC.²

Tencent Hy-Memory-30B-A3B official GGUF

Tencent published an official Apache-2.0 llama.cpp pack of Hy-Memory-30B-A3B. The card was created on 19 October at 07:52 UTC and last modified at 13:38 UTC. Q4_K_M-imat is 18.237 GB. Q6_K-imat is 24.707 GB. Q8_0 is 31.986 GB.³

Hy-Memory is a memory writer for Mem0 and MemOS, not a general chat model. The BF16 parent is still the 60 GB card from 16 October. Tencent lists LoCoMo Judge scores on the three GGUFs. Those scores are Tencent's.

TorchNeuron native neuron device

Hugging Face's official changelog now leads with TorchNeuron, dated 19 October. TorchNeuron is a PyTorch plugin that exposes AWS Trainium as a native neuron device. Transformers, Diffusers, and Kernels support it. Public beta starts on 20 October 2026. Install is pip install torch-neuronx.⁴

The post says new work should use those libraries directly instead of Optimum Neuron. Hugging Face claims 89 percent of the top 100 text-generation architectures work out of the box, versus 46 percent with Optimum Neuron v0.4.5. The worked example is Qwen3.8-27B, an existing card. This is a library change, not a new Hub checkpoint.

Daily Papers flip to Oct 20

Daily Papers flipped to an October 20 list. Fleet Experience (arXiv 2610.25736) leads with 87 upvotes. It is a collective learning recipe for robot fleets. There is a project page and no new Hub model card.⁵

OpenVLHarness from UCLA sits at 60 upvotes, with a project page and no Hub card.⁶ ScaleNav sits at 44. Its Hub card dates to 15 October.⁷ CarryOn is off today's Daily Papers list of 50.