Hugging Face Blog · 3h #

HiRO-ACEv2 simulates extreme weather at 3 km resolution cheaply

HiRO-ACEv2 is an AI weather model that simulates extreme weather at 3 km resolution for a fraction of the usual cost, released on Hugging Face.

Reddit · 4h #

Ling-3.1-flash: 560B-parameter MoE model with 1M context

A huge open MoE with 1M context pushes long-context open models further.

An open Hugging Face model with 560B total parameters, 25B active per token and a 1M-token context window. It follows Ling 3.0, which was 127B.

HF Papers · 1d #

Agent-Reflex trains language agents to reflect and recover from failure

Better reflection and recovery training could make deployed agents more robust in real tool environments.

Agent-Reflex is a training framework that teaches language agents to reflect on noisy MCP tool feedback and user corrections, improving failure recovery in long-horizon tasks.

HF Papers · 1d #

GreenWAM world foundation model targets robotics

A single world model serving prediction and control across bodies could simplify robot learning.

GreenWAM is a world foundation model that both predicts how a robot's actions change a scene and chooses those actions, aiming to work across different robot bodies.

HF Papers · 1d #

Tev scales long-horizon agentic environments with dense rewards

Dense reward environments could unlock open models for long-horizon autonomous tasks.

Paper Tev presents a method to scale long-horizon agentic training environments using dense progress rewards, aiming to help open models work autonomously for hours to days.

Reddit · 1d #

Texture-fix VAE for Qwen Image 2.1 released on Hugging Face

A VAE that reduces checkerboard artifacts in Qwen Image 2.1 generations, shared on Hugging Face. Differences are only visible when zooming past 500%.

Reddit · 1d #

Aziib Pixel Style LoRA released for Krea2 and Anima

A pixel-art LoRA for image models Krea2 and Anima, downloadable from Civitai with the trigger word 'aziib pixel style'.

HF Papers · 1d #

Urbex paper adds agentic spatial grounding for city-scale 3D

Urbex is a research framework that lets a vision-language agent actively search a city-scale 3D scene, querying landmarks and zooming instead of ranking precomputed instance proposals.

HF Papers · 1d #

VideoJev casts video understanding as decision tasks

VideoJev is a family of models that turn video understanding and interaction into structured decision tasks, adapting pretrained MLLMs with a shared prediction head instead of autoregressive text answers.

HF Papers · 1d #

Jeva proposes System One models from first principles

Jeva is a paper arguing current system-one model designs miss four structural requirements, including distinguishing answer absence from uncertainty and permutation invariance, and proposes models built around them.