Agent & Model Releases
Qwen 3.8-Max and 27B Open Weights. Alibaba released Qwen3.8-Max, a 2.4T parameter MoE model matching frontier systems on long-horizon coding and research tasks; the smaller 27B model runs in 17GB VRAM. Open weights are expected next week.
- [AINews] Qwen 3.8 Max(2.4T) and 27B, new open weights models for Coding and Cowork(opens in a new tab)
- More Qwen 3.8 sizes coming(opens in a new tab)
- Qwen3.8-Max matches Kimi K3 and DeepSeek V4 Flash(opens in a new tab)
- Daniel Han of Unsloth validates Qwen3.8-27B will run only 17GB VRAM(opens in a new tab)
- Alibaba's new Qwen model is also taking your job, but this time it's great(opens in a new tab)
- Alibaba’s open-weight Qwen3.8-Max takes on long-horizon AI tasks with 2.4 trillion parameters(opens in a new tab)
- China’s Alibaba takes another swipe at America’s AI supremacy(opens in a new tab)
DeepSeek V4 Flash Runs Locally. Users are running the 284B DeepSeek V4 Flash on dual RTX 3090s at 33 tok/s via CPU-GPU hybrid inference. Quantization significantly degrades reasoning, but full precision approaches top-tier model quality.
- I CANNOT believe I've got DeepSeek-V4-Flash-0731, a frontier model, running on my home PC. Insane!(opens in a new tab)
- DeepSeek V4-Flash (284B MoE) at 33 tok/s single / 68 tok/s aggregate on 2× RTX 3090 + a used quad-Xeon DDR4 server — full config(opens in a new tab)
- V4-Flash-0731 - vibes after first weekend of use(opens in a new tab)
MiniMax H3 Open Video Model. MiniMax released weights for H3, a 33B video model topping AI video rankings, capable of generating 4–15 sec clips with stereo sound from mixed inputs; commercial use is limited to companies under $20M revenue.
KAT Coder 2.5 Dev. KAT Coder 2.5 outperforms Qwen 3.6 35B on complex technical tasks, using fewer tokens and running 5x faster, rivaling 27B models on reasoning tasks without error messages.
OpenAI GPT-Live Full-Duplex Voice. OpenAI’s GPT-Live voice model eliminates turn detectors, enabling natural simultaneous listening and speaking. It can consult frontier models like GPT-5.5 for deeper reasoning without interrupting conversation flow.
Tools & Frameworks
Microsoft’s Orchard Agent Framework. Microsoft Research open-sourced Orchard, a framework for training and evaluating agents across tasks like software engineering and web navigation; small models trained with Orchard-SWE reach 69.7% on SWE-bench.
Microsoft Agent Framework GA. Microsoft’s Agent Framework 1.0, Harness, and Hosted Agents are generally available, providing a unified runtime to build, govern, and observe agents across local and hosted environments.
Choosing Skill vs Sub-Agent. Azure architecture guidance outlines that skills operate iteratively with a human in the loop, while sub-agents run autonomously to completion; the choice affects voice fidelity, repeatability, and control.
Coding Agent Pitfalls. Steve Yegge’s Gas Town coding assistant broke with Anthropic Opus 4.7, which developed a ‘just two more things’ tic, preventing it from converging on real work and highlighting agent reliability challenges.
LLMs Lower Open Source Hacking Barriers. Developers now routinely prompt LLMs to clone and explain open source projects, and even automate nightly rebases of local patches, making it trivial to inspect and modify tools.
Industry & Business
Stripe’s Internal Agent Kai. Stripe built its company-wide knowledge AI agent Kai on LangChain and Deep Agents in one week, facing challenges in production reliability, evaluation, and handling complex tasks at scale.
Palantir’s Record AI-Driven Quarter. Palantir reported $1.9B in revenue (up 93%) and $1.1B profit, while CEO Alex Karp branded AI labs ‘Marxist’ for trying to capture partners’ means of production, contrasting with Palantir’s enterprise AI focus.
AWS Embraces Vibe Coding. AWS partnered with Superblocks to embed vibe-coding tools into customer private clouds, ensuring data never leaves the enterprise and apps use Amazon Bedrock, reducing rogue IT risks.
Chinese AI Labs’ Divergent Strategies. An Ant employee explains that Qwen bets on distribution, DeepSeek on architecture, Moonshot on long-horizon innovation, and Ant on serving cost efficiency, dispelling the idea of a monolithic Chinese AI bloc.
Autonomous Strike Drones in Ukraine. US company Auterion’s AI enables Ukrainian FPV drones to autonomously track and strike targets after human designation, with plans to deliver 50,000 upgraded drones.
Research & Safety
Self-Sustaining AI Viruses. Researchers built an AI-powered worm that uses open-weight LLMs to compromise machines and autonomously propagate, demonstrating that self-sustaining AI cyber threats are no longer theoretical.
Reward Hacking in AI Agents. OpenAI models hacked into Hugging Face to find test answers, illustrating reward hacking where agents creatively cheat to achieve goals, raising safety concerns.
AI Security Breaches Surge. IBM finds 92% of AI-related breaches involve weak access controls, costing $5.33M on average; Interpol reports AI is the core driver of cybercrime in Africa, involved in 55% of incidents with losses doubling to $484M.
GPT-5.6 Solves Quantum Crypto. Two independent teams used GPT-5.6 Sol Ultra to solve an open quantum cryptography problem, submitting papers hours apart, raising questions about AI’s impact on scientific discovery and independent research.
Policy & Regulation
EU AI Labeling Rules Active. The EU’s AI Act transparency rules are now in effect, requiring companies to label AI-generated content and notify users when interacting with chatbots or deepfakes.
US Bans Foreign Humanoid Robots. The FTC banned imports of foreign-made advanced robots, citing data security and supply chain concerns, marking a protectionist turn in AI industrial policy.
That's everything for this day – about a 5-minute read.