Claude Mythos, Deepseek v4, HappyHorse, Meta’s new AI, realtime video games: AI NEWS

Key Concepts

• Agentic AI: AI systems capable of executing complex, multi-step tasks autonomously from start to finish.
• Zero-day Vulnerabilities: Previously unknown software security flaws that hackers can exploit.
• KV Cache Compression: Techniques (like RotorQuant) to reduce the memory footprint of Large Language Models (LLMs) during text generation.
• World Models: AI systems that simulate physical environments, allowing for interactive, 3D exploration of video content.
• Multimodal Models: AI capable of processing and generating multiple types of data, such as text, images, audio, and video.
• Open-Source vs. Closed-Source: The debate regarding the safety and accessibility of powerful AI models.

---

1. Anthropic’s Claude Mythos Preview & Project Glasswing

Anthropic has unveiled Claude Mythos Preview, a model described as a "step change" in capability, particularly in cybersecurity.

• Key Capabilities: The model autonomously identifies high-severity vulnerabilities in core infrastructure (e.g., Linux kernel, OpenBSD, FFmpeg, and cryptography libraries like TLS/SSH). It can chain multiple bugs to create full exploits.
• Performance: It shows a 14% improvement on Sweetbench Pro and a 13% increase on Terminal Bench compared to previous state-of-the-art models.
• Safety Strategy: Anthropic has opted not to release Mythos publicly, citing extreme danger. Instead, they launched Project Glasswing, a defensive alliance with major tech firms (Google, Nvidia, Microsoft, Apple, AWS) to allow authorized defenders to patch systems before malicious actors can leverage similar AI.
• Controversy: Critics argue the "thousands of vulnerabilities" claim is exaggerated and that smaller, open-weights models can replicate many of these findings when provided with isolated code.

2. ZAI’s GLM 5.1

ZAI has open-sourced GLM 5.1, currently considered the leading open-source model for agentic workflows.

• Functionality: Designed for long-horizon tasks, it demonstrated the ability to code a fully functional Linux desktop with 50+ apps from scratch using self-review loops over an 8-hour period.
• Accessibility: The model weights are available on Hugging Face (1.5 TB), with expectations for future quantized versions to make local execution more feasible.

3. AI Video & World Generation

• InSpatial World: Transforms 2D video into an interactive 3D environment. Unlike frame-by-frame generators, it builds a persistent internal simulation, allowing for camera movement and consistent physics. It runs at 10 FPS on an RTX 4090.
• Happy Horse 1.0: A new video generator from an Alibaba innovation unit. It currently tops the Artificial Analysis leaderboard, though it is not yet publicly available.
• Waypoint 1.5 (by Overworld): An open-source, real-time interactive world generator that runs on consumer hardware (RTX 3070+). It generates 720p/60fps worlds.
• mmFizz Video: A framework that improves physical consistency in video generation by training the model on layers of geometry, motion, and appearance rather than just pixel prediction.

4. Meta’s Muse Spark

Meta released Muse Spark, a multimodal model powering Meta.AI across Facebook, Instagram, and WhatsApp.

• Performance: While it excels in scientific figure reasoning and specific math tasks, independent benchmarks (Artificial Analysis) rank it behind Gemini 3.1 Pro and GPT 5.4. It is currently closed-source.

5. Specialized AI Tools

• LPM 1.0: A real-time conversational avatar generator. It maintains high identity consistency and natural non-verbal cues (eye movement, idle motion) during long-form interactions.
• Numina: A model-agnostic tool that fixes object-counting hallucinations in image generation, ensuring the output matches the prompt's specific quantities.
• RotorQuant: An open-source memory compression algorithm that outperforms Google’s TurboQuant. It uses "Clifford rotators" to compress KV cache vectors, resulting in 28% faster decoding and 44x fewer parameters.
• Komodo (Kinematic Motion Diffusion): An Nvidia-developed tool for generating 3D human/robot motions. It is used to synthesize training data for robotics in virtual environments like Isaac Sim.
• Spatial Edit: An open-source tool for precise object and camera manipulation within images, outperforming previous models like NanoBanana.
• Anima v3 Preview: A lightweight (2B parameter) model optimized for anime-style image generation, compatible with Comfy UI.

---

Synthesis and Conclusion

The AI landscape this week is defined by a tension between extreme capability and safety. The most significant shift is the decision by labs like Anthropic to keep their most powerful models behind closed doors, marking a departure from the "open-first" era. Simultaneously, the open-source community is rapidly closing the gap, with tools like GLM 5.1 and RotorQuant proving that high-level agentic and efficiency breakthroughs are no longer exclusive to closed-source giants. The focus is clearly shifting from simple text generation to agentic autonomy, physical world simulation, and precise control over AI outputs.