ThursdAI - The top AI news from the past week · From Weights & Biases, Join AI Evangelist Alex Volkov and a panel of experts to cover everything important that happened in the world of AI from the past week

🦃 ThursdAI - Thanksgiving special 25’ - Claude 4.5, Flux 2 & Z-image vs 🍌, MCP gets Apps + New DeepSeek!?

·1 hr 21 min·5 clips

Prime Intellect released Intellect 3, a 106-billion-parameter MoE with 21 billion active parameters.

1. ThursdAI’s Thanksgiving special on November 27, 2025 focuses on the week’s AI releases and the show’s third Thanksgiving episode in a row. 2. Alex Volkov hosts with Wolfram Ravenwolf and Jan Peleg, both described on-air as AI evangelists, and they matter because each brings a different release to highlight. 3. The episode exists to give a fast TLDR of major AI news, then slow down for deeper discussion, demos, and product comparisons. 4. Anthropic’s Claude Opus 4.5 is presented as the headline model, with the hosts saying it is their “premier intelligence” and a state-of-the-art coding model. 5. The show says Opus 4.5 reaches 80.9 on SWE-bench Verified, and the hosts compare that score to GPT-5.1 and Gemini 3 Pro. 6. The discussion also mentions Anthropic’s new effort controls and a very hard internal engineering test that Opus 4.5 reportedly beats. 7. MCP gets a one-year anniversary segment, and the hosts say the protocol has now been adopted by Google, OpenAI, Microsoft, and Anthropic. 8. The guests say MCP still lacks UI, so a new effort called MCP UI is being standardized to give apps a shared interface layer. 9. The hosts note that the MCP UI effort comes from Ido and Liad, who were previously on the show for Git MCP. 10. Open-source coverage starts with Prime Intellect’s Intellect 3, which the hosts describe as a 106-billion-parameter MoE with 21 billion active parameters. 11. The Prime Intellect segment includes benchmark claims such as about 90% on AI2ME 2025 and strong results on coding and reasoning tasks. 12. The hosts say Prime Intellect published unusually detailed information about post-training, infrastructure, and verifiers, which they treat as part of the release value. 13. DeepSeek Math V2 is introduced through a Clem Delang post, with the hosts noting Apache 2 licensing and a claimed IMO 2025 gold-medalist connection. 14. The DeepSeek discussion turns into a limitation analysis when the hosts point out the 168,000-token context window and how that can be too small for some proofs. 15. Tencent’s Hunyuan OCR is described as a 1-billion-parameter model that beats larger OCR systems on several benchmarks, including OCR Bench and document parsing. 16. Microsoft’s Pharah is framed as a local computer-use agent with 73.5 on WebVoyager and under-500-millisecond latency on Copilot Plus PCs. 17. The hosts spend substantial time on Flux 2, describing its 32-billion-parameter architecture, 10-image reference editing, and production-vs-open-release tradeoffs. 18. The episode’s tone is fast, technical, and demo-heavy, with the hosts jumping between model specs, benchmark numbers, product licensing, and live examples. 19. People interested in open-weight AI models, coding benchmarks, and multimodal releases would get the most from this episode. 20. People wanting a calm, non-technical holiday conversation probably should skip it.

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from ThursdAI - The top AI news from the past week