Today, we’re announcing Ternary Bonsai 2 27B. Based on Qwen3.8 27B, Bonsai 2 27B is 9x smaller than its full-precision counterpart while retaining 98.2% of its aggregate benchmark performance. Two months after the first Bonsai 27B release, the biggest change is quality. The footprint remains 5.9 GB, but the gap to full precision has narrowed materially, with particularly strong gains in agentic coding, multimodal reasoning, and long-horizon tool use. Ternary Bonsai 2 27B is available today under Apache 2.0.

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x cheaper (w/ output tokens free) • Frontier composable intelligence optimized for decisions AFAICT the shortest path to AI-based economic revolution

WARNING: I’m doing something a little crazy on Friday. & it might make a lot of people a LOT of money. I’m giving away $1,000,000 in cash and prizes live, in front of 200,000 people. Why do something this reckless? I’m launching my new book on scaling businesses. And I don’t want to do another predictable book launch where we count how many copies we sell. I want people to walk away with actual money in their pockets, and the tools to make even more. So here’s what’s happening: → We’ll give $25k business grants to four owners in the crowd → We’ll give away over a million dollars in prizes designed to help you run and grow your business → We’ll have special guests who’ve built and invested in billion-dollar companies join me on stage to share their playbooks with you Oh, and I’m also giving away my best frameworks for scaling businesses: → The 12 profit levers you can pull to grow your bottom line today → How to find and sell clients who pay more, stay longer, and are easier to work with → How to hire A-players who move your business forward → How to build a team that can run without you → How to increase your margins and stabilize your profits I’ve never put this much time, energy, or money into a free event before. It could completely implode. Or it could become the best book launch ever. Either way, it’s going to be crazy. Save your free spot at the link in my bio.

This will be a bigger deal than a new model release. It's called Function Hooks, branded as Claude Mods. Here's what it actually means. Today, you can extend Claude Code with basic hooks. Small shell scripts that run at fixed moments. Before a tool runs. After a session starts. That's it. They are limited, and on Windows they often break in weird ways. Function Hooks change that completely. Now a plugin is a real function that can sit inside almost anything Claude Code does. It can watch a tool call, block it, rewrite it, or let it pass through to the next plugin. Think of it like middleware, the same pattern used in web servers like Express. This unlocks real things: – Hide secrets from the AI before it ever reads them – Block a dangerous action for real, not just hope the AI follows a written rule – Build actual interfaces (someone already made Tetris) – Let admins control exactly what plugins are allowed to touch The comment section on the GitHub issue is not hype. It's a live audit. Engineers grabbed the early build and started trying to break it on purpose: – They made a hook hang, to see what happens on timeout. It failed silently, and the risky action ran anyway – They blocked a file write. The AI just picked a different tool and wrote it anyway, because the block gave it no reason. – They ran the safety checker on broken code. It said "all good" when it wasn't. This is a group of production users doing security testing for free, before the feature even ships, because they plan to run this in real systems with real stakes.

China published the most uncomfortable paper on vibe coding. ETH Zurich tested 100 developers in a controlled, commercial-grade vibe coding environment to see who actually succeeds. The findings are brutal. The researchers tracked computer science achievement, written communication skills, and general cognitive reasoning. They wanted to see what actually predicts vibe coding proficiency when you never touch a line of source code yourself. Two major predictors emerged. Written communication proficiency mattered. The ability to structure thoughts and articulate intent unambiguously in text directly impacts what the AI builds. But that wasn't even the main takeaway. Computer science achievement was a massive, dominant predictor of success. Even when researchers controlled for general intelligence and reasoning skills, CS background still heavily dictated who built working software and who completely crashed. In fact, CS knowledge contributed roughly twice the unique predictive variance of writing skills alone. Why? Because vibe coding isn't about writing code. It’s about debugging logic. When an AI agent builds a complex application and quietly breaks an edge case under the hood, a non-technical user looks at the glowing UI and assumes it works. They don't know what questions to ask. They don't know what logic to challenge. They lack the mental models to recognize architectural catastrophe. You can prompt your way past syntax. You cannot prompt your way past a fundamental lack of engineering intuition. The hype told us that learning to code is dead because language is all you need. The data just proved the opposite. To truly master the vibe, you still need to understand how the machine thinks.

“The core technology is MiniMax H3's accelerated version H3 Max, with post-training + inference optimization by @fal: 1. About 5-second video generated in under 3 seconds 2. 15-second video output in about 9 seconds 3. Throughput an order of magnitude higher than official H3 This means AIGC is no longer just produce-first-then-post; it can now enter live streams in real time”