Tencent open-sources 770B-parameter Hy4 preview with 1M-plus-token context

Hy4 is already processing trillions of tokens on OpenRouter within a few days, apparently more than GLM 5.3 did in a week. Its 5% cache-read cost is especially compelling when 10% to 20% is common.

I would not trust OpenRouter usage figures blindly. Publishers may be paying for exposure or buying their own tokens, as the rest of the AI business does.

Cache reads are often the biggest hidden API cost. I still have not found a decent model below $0.01 per million cache tokens except DeepSeek's official API, even after its price increases.

I used Hy3 quite a bit and liked it for the tasks it suited, but it was slow through Tencent and other inference providers. A smaller model should have been served faster.

My Hy3 evaluations saw an unusually large jump from preview, where it was unusable, to the final release, where it was competitive. I wonder whether Hy4 will change that much too.

The bar charts in Tencent's post look strange to me. The displayed bar heights do not seem to match the numbers consistently.

I think the first chart is visually confusing rather than wrong: Hy4 and Hy3 are overlaid in the first column, with Hy4 as the taller dark-blue bar and each score printed near its own bar top.

I tried Hy3 on a project and was surprised by how good it was as a general-purpose agentic model. In our tests, only DeepSeek4-flash beat it, and its behavior was close enough to DeepSeek that I wondered if it was forked.

Current models are adequate for nearly everything I want, but we have saturated our ability to verify their output at scale. More autonomy makes intent and correctness harder for humans to check, and I think that needs a different architecture with durable, plastic memory.

Hardware debugging and firmware work still create long think-test loops where frontier models fall short.

Even Opus 5 writes buggy, low-quality code and makes serious mistakes in code analysis. Frontier-model swarms have not produced a working browser or compiler-scale codebase, and reverse engineering remains hit-or-miss.

I want to generate a full movie from a book, animated or live action. Existing models cannot maintain context across tens of thousands of frames.

I use a scrambled picture cube without a reference image as an AGI test: I give models videos of it and ask them to solve it. None has solved it yet.

I asked ChatGPT, Claude, and Gemini to identify a laptop from its brand, screen size, and ports. Claude and ChatGPT got there only after multiple rounds and wrong guesses, while Gemini never found the exact match.

Model capability improvements shift the whole price-performance curve, like decades of computer hardware upgrades did. Today's budget models already beat older premium models, so better frontier models can improve everyday use too.