This benchmark tests vision-language models on generating and editing ASCII art in terminal environments, providing a new evaluation task.
A developer released a two-stage prompt-injection detector combining a regex gate with a quantised DeBERTa-v3 model in ONNX format, trained on real attack data from a game.
This paper presents Audio-Visual Flamingo, a model architecture designed to understand long and complex videos by integrating audio and visual modalities.
Intern-S1 is a project from the InternLM team at Shanghai AI Laboratory, gathering 834 stars. The exact nature is unclear but likely a model or tool release.
Qwen-AgentWorld is a new agent framework released by the Qwen team at Alibaba, with 850 stars on GitHub. It likely supports building and managing AI agents.
This paper investigates the effectiveness of the Muon optimizer in agentic reinforcement learning tasks, providing guidelines on when it outperforms standard optimizers.
This paper proposes expanded hyper-connections to widen residual streams in neural networks, aiming to improve model intelligence and scaling properties. It was shared on r/LocalLLaMA.
Kimi CLI is a command-line interface for interacting with Kimi, a large language model from Moonshot AI. It has over 10,000 stars on GitHub as of July 20, 2026.
This study analyzes how reasoning capabilities emerge and evolve in LLMs across pretraining and post-training phases. It received 12 upvotes on Hugging Face Papers.
This paper introduces PagedWeight, a technique for efficient serving of Mixture-of-Experts LLMs using dynamic quantization that balances quality and throughput.
Outerport, a YC S24 startup, launched a service that enables instant hot-swapping of AI model weights without downtime. The launch received 93 points on Hacker News.