Tencent Open-Sourced a 770B Model That Helped Optimize Itself

Somewhere in Tencent's press release on Friday, between the pricing table and the free-trial fine print, there's a paragraph that should have been the entire announcement.

It says that Hy4 preview — Tencent's new flagship open model — "proposed approaches, ran experiments, and iterated based on the results" while optimizing its own inference infrastructure. Then it casually mentions the model found wins in operator fusion and communication optimization that added up to 31.8% more end-to-end throughput.

Read that again. A model sped up the machines it runs on, by itself, and the news was delivered in paragraph nine of a product announcement, right next to a note about free trials.

If a lab in San Francisco had shipped that sentence, you'd have seen a keynote, forty X threads, three "what this means for AGI" essays, and at least one panicked op-ed. Tencent put it in a press release with the same energy companies use to announce a new pricing tier.

So let's actually talk about what shipped, because the rest of the release is arguably just as interesting — and a lot more useful to you if you build things with models.

The spec sheet, minus the confetti

Hy4 preview is a mixture-of-experts model: 770 billion total parameters, 49 billion active per token. Context window is a hair over a million tokens (1,048,576, if you like round numbers that aren't round). And unlike a lot of "open" releases, it's genuinely open source — Apache 2.0, up on Hugging Face, weights and all.

Pricing through the API: $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million tokens for cache hits.

That last number deserves more attention than it's getting. Cache reads at roughly 5% of the input price is aggressive — the industry norm hovers closer to 10–20%, and prompt caching is the single biggest cost lever an agent has. Agents re-read the same files, the same system prompts, the same tool outputs, thousands of times an hour. If you're running an agent that spends 80% of its tokens on repeated context, Hy4's cache pricing quietly cuts your bill in a way no benchmark score will ever show you.

The MoE math matters too. You download (or rent) a 770B-parameter model, but every token only activates 49B of it. That's the difference between "needs a data center" and "needs a well-configured inference cluster." It's the same economic trick DeepSeek pulled, refined again: frontier-ish quality at a fraction of the serving cost.

The self-improvement part, honestly

Let's not get carried away, because the release certainly doesn't. Tencent calls it an "early-stage recursive self-improvement loop," which is carefully hedged language for something real but narrow.

What actually happened: the model was pointed at the unglamorous plumbing of AI infrastructure — fusing operators, optimizing how GPUs talk to each other — and it found improvements that compounded to 31.8% throughput. It proposed experiments, ran them, looked at the logs, and iterated. The code, logs, and feedback fed the next round.

Is that AGI? No. Compiler autotuning by a very well-read model with tool access is exactly the kind of task current models are good at: huge search space, clear success metric, no ambiguity about whether the thing worked. You could argue this is just automated systems engineering with good PR.

But here's the part that should make you a little uneasy in a productive way: the loop is the product now. The model that ships is the model that improved the next shipping model. Tencent didn't describe a research demo — they described a step in their training pipeline that happens to be made of the model itself. Every serious lab is attempting some version of this. Tencent is the first to fold it into a Tuesday press release like it's a feature bullet.

The download-a-credit-card moment isn't "AI got smarter." It's "AI improving AI stopped being a paper and became infrastructure."

The benchmarks, with a shaker of salt

The headline eval is a blind test Tencent ran internally: 163 experts scoring 203 real engineering tasks. Hy4 preview averaged 2.99 out of 4.00, nudging out Kimi K3 (2.94) and GLM-5.3 (2.92).

Some perspective on those numbers. First, it's an internal eval — Tencent picked the tasks, the raters, and the rubric. Second, the margins are thin enough that you could re-run this next week and get a different podium. Third, "internal blind evaluation by the vendor" is the benchmark equivalent of grading your own homework, even when the grading is genuinely blind.

What I'd actually trust more is the shape of the claim rather than the digits: Hy4 is in the top tier of open models, clustered with Kimi K3 and GLM-5.3 rather than clearly above them, and the standout areas are long-context software engineering, office-document workflows, and — interestingly — scientific domains like molecular dynamics and condensed-matter physics. Also game dev: the release claims it can generate a playable prototype from a single natural-language request, which I have not verified but really want to.

One more data point from the community that I find more predictive than any leaderboard: developers who tried Hy3's preview last cycle reported the gap between preview and final release was "unprecedented — from unusable to competitive." Previews from the Hunyuan team are drafts, not finished products. Judge Hy4's final release when it lands; meanwhile, treat the preview as a promising beta that happens to be cheap.

Actually open source this time

Every time a Chinese lab ships weights, a certain corner of the internet shows up to say "open weights aren't open source" — no training data, no recipe, so no real openness. It's a fair debate, and I'm sympathetic to it.

But the license matters, and Apache 2.0 matters a lot. You can take Hy4, fine-tune it, fold it into a commercial product, run it on your own hardware, and never send Tencent a token or a cent. Whatever purists say about reproducibility, that's more practical freedom than most "open" models from American companies offer — many of which come with usage restrictions, output mandates, or terms that lawyer-speak into oblivion the moment your traffic gets interesting.

The open-weights crowd has spent two years proving the uncomfortable point: open models don't need to beat the closed frontier. They need to be good enough and 10x cheaper, and the market does the rest. Hy4 at $0.834/M input with 5% cache reads is the latest exhibit.

About that OpenRouter traction

The claim making the rounds: Hy4 pulled trillions of tokens through OpenRouter within days — reportedly more volume than GLM-5.3 did in its entire first week. Those specific numbers come from community tracking, not an audited dashboard, so hold them loosely.

Hold them especially loosely because the skepticism writes itself. A company with Tencent's distribution could plausibly juice router stats — route internal traffic through the public endpoint, subsidize a launch week, whatever. One HN commenter put the suspicion bluntly: "If you're Tencent you can just plug it into some field somewhere that lots of people see."

But here's the thing — even if you haircut the numbers by half, twice, the direction is unmistakable. Every few months a Chinese open model lands on OpenRouter and immediately does numbers that frontier closed models spent years reaching. The default developer playbook in 2026 increasingly reads: prototype on the cheap open model, escalate to the closed frontier only when you hit a wall. A few years ago that sentence would have been laughed off HN. Now it's just how people ship.

Chart crimes and other human details

No release post in 2026 is complete without the benchmarks-chart forensics, and the HN thread did not disappoint. Multiple people noticed the bar charts in Tencent's announcement didn't quite match their own numbers — bars whose heights didn't correspond to the scores printed under them. Probably an overlay-rendering quirk rather than malice, but "top-tier AI lab ships visualization that fails its own benchmark" is a very specific kind of comedy.

My favorite thread, though, was people sharing the model's reasoning traces from a drawing task. Hy4 deliberating over whether a pelican should wear a tiny helmet — "maybe add a helmet? could improve riding theme... but a helmet might obscure" — is either the cutest thing you'll read this week or a reminder that these systems hold opinions about tiny hats. Both, honestly.

The point of including this: the discourse around these releases oscillates between "civilizational threat" and "stochastic parrot," and the truth is a model that speeds up its own GPU cluster by a third and then fusses over pelican headwear. Weird machine. Getting weirder.

What this actually means for you

If you build with agents, three practical takeaways:

The price floor keeps falling, and cache pricing is the new battleground. Everyone compares per-token headline prices, but your agent's real bill lives in cache hits and repeated context. A model that underprices cache reads will beat a "cheaper" model in production every time, because agents are repetition machines. Watch the cache column, not the input column.

Million-token context is becoming table stakes, and it changes what you build. When an agent can hold your entire codebase — not a summarized, RAG-chunked ghost of it, the whole thing — entire categories of retrieval plumbing become optional. The models are now ahead of the harnesses. That's where the interesting engineering is.

And the open-model ladder keeps climbing. The gap between "open tier" and "closed frontier" is now small enough that for most real work — coding, document workflows, data analysis — it's a pricing decision, not a capability decision. The interesting question for the closed labs isn't whether Hy4 beats GPT on some benchmark. It's why anyone would pay several times more for the privilege when the open alternative handles their workload fine.

Meanwhile, somewhere in a Tencent data center, a model is probably fusing another operator. The loop, as they say, is early-stage.

---

If you'd rather have an agent that puts models like this to work — browsing, running commands, reading your files — that's what we built CopperRiver for. Desktop AI assistant, open models, plans from $9/mo.