# 📆 ThursdAI - GLM 5, MiniMax 2.5, Seedance 2, Gemini 3 Deep Think, Codex Spark 100tps & more AI news Source: https://www.youtube.com/watch?v=wQb4JK5xKMw Recap page: https://rapidrecap.app/video/wQb4JK5xKMw Generated: 2026-02-13T06:30:58.244+00:00 --- ## Quick Overview The ThursdAI episode focused heavily on the explosive return of open-source LLMs, highlighted by the simultaneous releases of GLM 5 from Z.ai and Minimax 2.5, both models competing closely with frontier models like Claude Opus 4.5 and Gemini 3 Pro on key benchmarks, alongside the mind-bending video generation model Seedance 2 from ByteDance. **Key Points:** - GLM 5, released by Z.ai, claims the open-source coding crown and features 744 billion parameters with only 40 billion active, utilizing a DeepSeek sparse attention mechanism. - Minimax 2.5 achieved 80.2% SWE Bench verified, claiming state-of-the-art coding performance, and operates with only 10B active parameters, emphasizing speed and efficiency. - Lou from Z.ai stated GLM 5 is 'bigger, faster, better, and cheaper,' introducing a new asynchronous reinforcement learning framework called SLIM. - Olive Song from Minimax attributed their rapid scaling to their RL framework, Forge, focusing on training speed, stability, and balancing inference speed with tool-calling efficiency. - The panel expressed excitement over the cost performance ratio, noting Minimax 2.5 is significantly cheaper than Opus 4.6 for similar task completion rates. - Alex Volkov highlighted Seedance 2 from ByteDance as a personal excitement, noting its ability to shatter reality limits in character consistency, physics, and sound, though difficult to access without a Chinese phone number. - Minimax's team is highly AI native, with one colleague treating agents as interns to collect feedback and improve skills, demonstrating advanced internal AI utilization. **Context:** The ThursdAI episode hosted by Alex Volkov of Weights & Biases, alongside co-hosts Wolfram Ravenwolf and Ryan Carson, provided a rapid news update focused almost entirely on major breakthroughs in open-source AI models immediately following releases from major labs the previous week. The show featured live interviews with key researchers from Z.ai (Lou) regarding GLM 5 and Minimax (Olive Song) regarding their newly dropped Minimax 2.5, establishing a direct comparison between the leading open-weight models and proprietary frontier models. ## Detailed Analysis The main focus of the show was the dominance of open-source intelligence this week, featuring breaking news interviews with the teams behind GLM 5 and Minimax 2.5. GLM 5 from Z.ai showcases massive scale (744B parameters, 40B active) and claims superiority in coding benchmarks, beating Gemini 3 Pro on some evals, while emphasizing agentic engineering capabilities that allow the model to handle multi-step decomposition and tool calling effectively. Minimax 2.5, released just 30 minutes before the show, achieved an astonishing 80.2% on the difficult SWE Bench verified, leveraging only 10B active parameters and focusing heavily on efficiency, speed (100 tps mentioned for Codex Spark in the title context, though Minimax focused on task time reduction), and cost effectiveness via their Forge RL framework. Both models offer intelligence levels approaching or matching proprietary leaders like Opus 4.5/4.6 at dramatically lower inference costs, leading to discussions about the shift from renting intelligence to owning it, particularly relevant for agent orchestration via platforms like OpenCL. Beyond LLMs, Alex Volkov championed Seedance 2 from ByteDance, a video model so advanced it breaks reality limits in consistency and physics, although access remains restricted. ### Open Source LLM Showdown - GLM 5 vs Minimax 2.5: GLM 5 released yesterday with 744B parameters and DeepSeek sparse attention - Minimax 2.5 dropped today achieving 80.2% SWE Bench verified with only 10B active parameters - Both compete directly with Claude Opus 4.5 and Gemini 3 Pro on benchmarks. ### GLM 5 Innovations - Lou from Z.ai highlighted that the model is 'bigger, faster, better, and cheaper' due to increased scale and adoption of the SLIM asynchronous RL framework - The model performs well in agentic workflows, handling multi-step decomposition and tool calling effectively. ### Minimax 2.5 Development Focus - Olive Song detailed scaling success through their RL framework, Forge, prioritizing fast and stable training iterations - The model architecture remains similar to M2/M2.1 (10B active parameters), but performance gains come from RL focusing on reducing end-to-end task time and efficient tool calling. ### Performance and Cost Analysis - Minimax 2.5 showed a 57% win rate versus Opus 4.6 at 15 cents per task, compared to Opus 4.6 costing almost $3 per task - The speed and lower cost make these open models highly competitive for agentic use cases running on OpenCL. ### Agentic Engineering Insights - Lou confirmed the shift to agent engineering, noting GLM 5 builds real understanding of user habits over time - Olive Song described Minimax as a very AI native lab, where team members use agents as 'interns' to build the next generation models, though current models still struggle with complex RL development itself. ### Other AI News Highlights - ByteDance officially released Seedance 2, a video model with incredible character consistency and physics, supporting 9 images, 3 videos, and 3 audio clips as reference - Qwen launched Qwen Image 2, a 7B parameter image generation model near SOTA - Entire raised $60M seed funding for an open-source developer platform for AI agents.