## The big picture

- **NVIDIA Model-Optimizer** is a new unified library for state-of-the-art model compression and inference optimization, supporting quantization, distillation, and speculative decoding for deployment in TensorRT-LLM and vLLM. [GitHub](https://github.com/NVIDIA/Model-Optimizer)
- **Rufus-Air** releases a fully reproducible post-training recipe for a 106B model, detailing an eight-stage pipeline from SFT to RLHF that establishes a new benchmark for open-weight alignment without proprietary data. [HF Daily Papers](https://huggingface.co/papers/2609.29421)
- **AgentKernel** proposes a trust-native operating system for agents, moving security from application-level middleware to mandatory OS-level boundaries for identity, input mediation, and memory governance. [HF Daily Papers](https://huggingface.co/papers/2609.29647)

## Architectural breakthroughs

- **CLM-v0.1-8B** uses a contrastive learning objective with InfoNCE loss on a frozen Qwen3-8B encoder to achieve up to 9x lower latency for agentic tasks by caching state and action embeddings separately. [Hugging Face](https://huggingface.co/Contrastive-LM/CLM-v0.1-8B)
- **DeltaWAM** improves bimanual manipulation efficiency by predicting visual deltas rather than full frames, using Streaming Delta Memory to reduce heavy video-expert processing during inference. [HF Daily Papers](https://huggingface.co/papers/2609.28811)
- **ViRDM** eliminates the need for teacher-critic stacks in few-step causal video generation by post-training generators against precomputed target distributions via representation distribution matching. [HF Daily Papers](https://huggingface.co/papers/2609.28923)
- **Neural Spectral Capacity** introduces a closed-form scalar metric for architecture design based on the singular-value spectrum of weight matrices, allowing global optimization of depth-width trade-offs without model instantiation. [HF Daily Papers](https://huggingface.co/papers/2609.23087)

## New open-weight model releases

- **Audio8-ASR-Infinite** is a native streaming speech recognition model with a rolling KV cache that maintains constant memory and latency for unlimited-length transcription, featuring selectable audio clocks and semantic VAD. [Hugging Face](https://huggingface.co/Edge0/Audio8-ASR-Infinite)

## Hardware & optimization

## Also this week

- **vectorize-io/hindsight** — A trending GitHub repo addressing long-term memory in agents. [GitHub](https://github.com/vectorize-io/hindsight)
- **Training Object Permanence in World Models** — Evaluates video models on a 1.5M-sample dataset of cognitive science-inspired tasks. [HF Daily Papers](https://huggingface.co/papers/2609.28654)
- **How far behind Nvidia is Huawei?** — Epoch AI analysis suggests export controls make it unlikely Huawei will catch up by 2030. [Latent Space](https://epochai.substack.com/p/how-far-behind-nvidia-is-huawei)
- **Runway’s WorldPrompt** — Uses persistent context and timed actions to steer real-time world models. [Latent Space](https://www.latent.space/p/runway)
- **OmniEcho** — Introduces a benchmark for spatial audio-visual understanding in embodied agents. [HF Daily Papers](https://huggingface.co/papers/2609.23407)
- **Agent-Editing World Model** — Proposes editing noisy reasoning continuations to prevent task-state contamination in agents. [HF Daily Papers](https://huggingface.co/papers/2609.28416)
- **Parts-of-Speech as Emergent Categories in SAE Latent Space** — Shows PoS distinctions are recoverable from SAE activations but do not align with one-to-one latent mappings. [HF Daily Papers](https://huggingface.co/papers/2609.29362)
- **Google’s Project Suncatcher** — Plans to deploy ML infrastructure in space. [source](https://blog.google/innovation-and-ai/models-and-research/google-research/google-project-suncatcher-facts/)
- **Qwen-Planner-Agent** — A closed-loop AI-for-AI framework for mobile planner agents. [HF Daily Papers](https://huggingface.co/papers/2609.29892)
- **anthropics/skills** — Anthropic’s official release of agent skills for tool interaction. [GitHub](https://github.com/anthropics/skills)
- **IterSynth** — Decouples planning and synthesis in deep search agents to reduce context noise. [HF Daily Papers](https://huggingface.co/papers/2609.29444)
- **Self-Organizing Agent Teams** — Agents learn reusable strategies to organize roles and information flow dynamically. [HF Daily Papers](https://huggingface.co/papers/2609.22682)
- **Hackers influence ChatGPT and Gemini** — Reports of manipulation to direct users to scam centers. [source](https://medium.com/@arielsimon/dark-sourcery-how-hackers-manipulate-ai-to-scam-you-88df434d2073)
- **PUBG Ally** — A voice-enabled embodied agent that plays alongside humans in PUBG. [HF Daily Papers](https://huggingface.co/papers/2609.29837)
- **World Action Agent** — Uses VLMs for robot manipulation via world action rehearsal in a visual workspace. [HF Daily Papers](https://huggingface.co/papers/2609.29964)
- **ExplorationBench** — Measures AI exploration in verifiable alien worlds to distinguish discovery from recall. [HF Daily Papers](https://huggingface.co/papers/2609.30199)
- **The Linear Representation Hypothesis Needs a Group Action** — Formalizes representation equivalence using group actions for mechanistic interpretability. [HF Daily Papers](https://huggingface.co/papers/2609.27158)
- **Knowledge Pull Requests** — A framework for continual document authoring that separates knowledge changes from text edits. [HF Daily Papers](https://huggingface.co/papers/2609.26634)
- **Humans Are Reading Your ChatGPT Chats** — A lawsuit claims that OpenAI uses human review for user data. [source](https://openclassactions.com/lawsuits/privacy/openai-chatgpt-human-review-project-lily-class-action-lawsuit.php)
- **For Computer Use, the harness matters as much as the model** — [source](https://www.stagehand.dev/evals)
- **Oracle invokes force majeure** — Oracle has invoked force majeure regarding a New Mexico AI data center. [source](https://qz.com/oracle-force-majeure-new-mexico-ai-data-center-092426)
- **The AI Build-Out Is Becoming the Biggest Economic Bet in U.S. History** — Macro-economic context on AI investment. [source](https://www.wsj.com/economy/the-ai-build-out-is-becoming-the-biggest-economic-bet-in-u-s-history-c60716dd)