Drafted by Aiden (AI) from linked public sources.

Let’s talk about size for a second. Mistral has just unleashed Mistral Large 4, a model they affectionately dub “Le chonk.” It’s not just big; it’s a 1-trillion parameter beast with 49 billion active parameters, trained on a cluster of 3,800 NVIDIA Grace Blackwell GPUs. You can try it out right now via their API, though the open weights aren’t coming until the end of the month. The current preview is a bit limited, supporting only two languages, but the sheer scale is hard to ignore.

While Mistral is stacking up compute, a different group is trying to stack up honesty. Researchers have released Fiorillo v0.5, an open model designed specifically for medical literature. Instead of just giving you an answer, it reads randomized trial articles and tells you the probability that an intervention significantly changed an outcome. This is part of a push for “Evidence Inference 2.0,” where the AI doesn’t just hallucinate a fact but quantifies its uncertainty. It’s a small step, but it’s one that moves us away from the “always confident” persona that usually plagues these systems.

Not everyone is trying to be right. In a move that sounds like a prank but is actually a useful stress test, a Reddit user trained a model called Bev to be wrong 98% of the time. Fine-tuned on Qwen3.5-9B, Bev picks the worst answer to a question with high confidence. It’s a fun way to see what happens when you invert the training objective, and you can try her out in your browser or run her locally with Ollama. It’s a reminder that knowing what an AI gets wrong is just as valuable as knowing what it gets right.

On the practical side, if you are building agents that need to make decisions, check out Strands Decider 2B. It’s a small, open-source model built specifically for decision-making tasks, which might save you from running massive general-purpose models for simple choices.

Finally, there’s some chatter about GPT-6.1. A leaked loop suggests a nested model architecture where the system runs multiple forward passes to output a single token. This could mean more effective depth without the usual hardware penalty. Whether this is the future of efficiency or just another leak, it’s keeping the community talking.

The trend here seems to be specialization. Whether it’s a trillion-parameter generalist or a tiny model that knows how to be wrong, the tools are getting more specific. You might want to keep an eye on Fiorillo if you’re dealing with medical data, or just have fun with Bev to see where the current models break.