
MIT-licensed and optimized for text generation, summarization, rewriting, and chat. Model: huggingface.co/microsoft/Aio…

Tweeting interesting papers submitted at huggingface.co/papers. Submit your own at hf.co/papers/submit, and link models/datasets/demos to it!

MIT-licensed and optimized for text generation, summarization, rewriting, and chat. Model: huggingface.co/microsoft/Aio…

Microsoft just released Aion Instruct on Hugging Face A 2.3B small language model built for on-device AI in Windows, with a 32K context window and support for 8 languages.

Paper: paperswithcode.co/paper/2611… Models: huggingface.co/collections/g… Validation set: huggingface.co/datasets/gaot…

Multi-Branch Transformer Scales LLMs by branching, running full Transformer blocks in parallel instead of only making them wider or deeper, reaching lower validation loss at matched compute and up to 51.4% faster decoding.


AscendGYM A cost-aware kernel optimization benchmark for real-world workloads on Ascend NPUs, with 48 kernels across 3 difficulty levels and a new Speedup-Token Score (STS) that weighs speedups against token cost

Paper: paperswithcode.co/paper/2611… Model: huggingface.co/XingChen-AGI/…

Xing4: frontier agentic LLMs on Hugging Face By China Telecom, trained on 8,192 Ascend 910C cards. 285B-param model with agentic post-training via TeleGym. Open weights on the Hub.

Paper: paperswithcode.co/paper/2611… Project: wsp-no-transformer.github.io… Code: github.com/effl-lab/Weight-S…

LLMs write weights by hand – and they choose tables over Transformers Three frontier LLMs, given ten examples and a sandbox, compile plain-text facts into sparse concept tables, reaching up to 54.4% ARC accuracy with 1,800× fewer parameters than a Transformer – and our associative-memory model runs 17× faster.

Find it here: huggingface.co/tencent/Hy4-f… Great for everyday productivity tasks — decoding speeds past 150 tokens/s with speculative decoding built in.

Tencent just released Hy4 flash on Hugging Face A 220B MoE multimodal model with only 12B active per token and 1M context, delivering Claude Haiku-level satisfaction at a fraction of the cost.

Paper: paperswithcode.co/paper/2611… Dataset on Hugging Face: huggingface.co/datasets/Toke…

SWE-CoWork A new benchmark for coding agents as digital employees. Agents must recover requirements distributed across 26 simulated colleagues and workplace records, then implement and verify changes. Leading models pass ~63% of new-behavior tests but finish under 8% of full tasks.


Apple just released FastFrames-L on Hugging Face A 14B autoregressive video diffusion model that generates 480p video at interactive latency - under 50ms per frame.