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

📆 Open source just pulled up to Opus 4.6 — at 1/20th the price

·1 hr 28 min·5 clips

MiniMax 2.5 reaches 80.2% on SWE-bench Verified.

1. ThursdAI covers GLM5, MiniMax 2.5, C-Dance, and other AI releases from February 12. 2. Alex Volkov hosts with Wolf, Ryan Carson, Jan Pelik, and guests Lou from ZAI and Olive Song from MiniMax, all of whom speak from active release work. 3. The episode asks whether open-weights models are catching frontier labs on coding, agents, and price. 4. Alex opens with GLM5 and says the model is already live on Weights and Biases with day-zero support. 5. Lou says GLM5 uses a new asynchronous reinforcement learning framework called SLIM. 6. Lou also names DeepSeek sparse attention as part of the update and says it preserves long-context performance while reducing deployment cost. 7. The hosts compare GLM5 against Claude Opus 4.1 and Gemini 3 Pro on benchmark charts shown on screen. 8. Alex highlights Humanity's Last Exam, the terminal bench, and vending bench as the evaluations that matter most. 9. Lou says GLM5 is meant for agentic engineering, where the model decomposes multi-step tasks and uses tools on its own. 10. Lou says the model is better suited to memory-driven workflows because it can adapt to user habits over time. 11. The panel says GLM5 is 744 billion parameters with 40 billion active, which makes local use unrealistic for most people. 12. Wolfram says the English version impressed him more than the earlier German testing, which had felt weak. 13. Ryan Carson says the release matters because people want to run agentic workflows like OpenClaw while they sleep. 14. Jan Pelik says GLM5 looks close enough to Opus 4.5 that repeatability may matter more than raw benchmark score. 15. Alex notes that ZAI also released its own terminal bench work, which Lou says came from internal research on better evaluation. 16. Olive Song says MiniMax 2.5 was still training when the team chose to release it, because the improvement was already large enough. 17. Olive says MiniMax focused on reinforcement learning, task efficiency, and lower end-to-end task time rather than only benchmark scores. 18. The conversation with Olive is paced like a live launch interview, while the show overall keeps a fast news-roundup rhythm. 19. Listeners tracking open-weights coding models and agent tools will get the most value. 20. Viewers looking for a slow, single-topic deep dive may skip this one.

Listen to the show on

Apple Podcasts Spotify

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