Been testing @TencentHunyuan Hy4 preview in @WorkBuddyAI, and I came away pretty impressed with the document work. The game demos in the walkthrough were a lot of fun too. 770B parameters, 49B active, a 1M-token context window and Apache 2.0 weights. I wanted to see what that looks like in actual tasks. The Mario-style platformer felt really good to play. Holding jump changed the height, mushrooms popped out of blocks, and enemy stomping worked. The movement and collisions were the bits I wanted it to get right, and the demo came together nicely. The neon racer looked great, especially the wet-road lighting. Drifting into a boost widened the camera and kicked out blue-purple flames, while AI cars raced alongside. That was probably my favorite visual demo. The expense audit impressed me for a different reason. It worked through 24 claims, checked policy versions and flagged duplicate invoices and allowance issues. I could follow the evidence behind its decisions, which made the output useful. Even the research deck came out well: 10 slides, a comparison table, and numbers that matched my spot checks. I liked getting the research and presentation together. Two prompts to try, shortened from the demos: Mario: "Build a Mario-style platformer using HTML5 Canvas and vanilla JavaScript. No game engine. Use requestAnimationFrame, terrain and brick collisions, mushroom and coin blocks, enemy stomping, gravity, inertia, acceleration and jump physics. Draw the pixel art directly with Canvas instead of using images." Racing: "Create a nighttime neon street racer in Three.js with a third-person chase camera, wet reflective roads and arrow-key driving. Drifting builds nitrous. Boosting widens the camera and produces blue-purple exhaust flames. Include AI opponents, live standings and a three-lap timer." I'd highly recommend you to try Ponytail to simplify the game code, plus a frontend design skill to polish the racing HUD, typography and boost meter. I haven't tested that combination on these demos yet, but I'd like to see how much further it can take the results. Give @TencentAI_News Hy4 preview a spin in WorkBuddy and drop your builds in the replies. Curious what you'd throw at it.

@TencentHunyuan Hy4 preview is a really good and frontier level model and it's currently free on @WorkBuddy_AI. The model is good across the board from document work to coding work to fun apps creation. It's super cool to chat with as well! I did a full video testing it on YT, check it out:

This is very true. Yes, the classifiers have changed. Hence, it can fallback to Opus more often but that doesn't change the performance of Fable at all. It's just the routing that can be weird and lead to lower "un-supervised" and untrue benchmark results. Though, Sonnet 5 is bad.

GLM-5.2 on KingBench (3). Thoughts: The model has superb taste. It is greater at UX than UI. The code is always very clean. It is great at One-shot wonders. I asked it to fine-tune a whole local model and it did it in 30mins! This is just a great model to use all-round. 1/n