Herlon Aguiar on X: "As promised, here's @Cloudflare's new model "Clef" using NVFP4 quant playing @BalatroGame and winning at ante 8, fully offline on a single RTX 5090! The last play is mindblowing, very nice combo! I'm very impressed that it managed to beat the game. Balatro is not a simple game and this model doesn't have a CoT to plan things better like LLMs do. It's not a god-like model, it still makes some decisions that aren't very good but it managed to beat the game in the end, so ¯\_(ツ)_/¯ The response time is not fast, almost 4s on average. The larger the context size the slower it gets. Total playthrough duration was 19m 40s (~3min faster than GLM 5.3 Flash NVFP4 in the sparks). I had to get close to the 16K token limit trying to squeeze as much info as I could while also attaching a lower resolution screenshot of the game (640x426) thanks to the vision encoder in the model 🙌 This one was especially tricky, as I had to make the options clear and try not to bias and make it an agreeable machine by adding options that I would consider "bad". I can't say that it decided well if there were no bad options to choose from. The economy seems very different from the GLM 5.3 Flash playthrough, I might give GLM another run with the latest state of the harness to see if it will also spend more money instead of just keeping it. Also it used a lot of discards, which was quite interesting. I still think there are a few issues with the harness and I will keep improving it as I try more models. @CloudflareDev"