Tencent has quietly shipped Hy3, the full release of its Hunyuan-line flagship, and the numbers make it hard to ignore. It is a 295-billion-parameter mixture-of-experts model that only activates 21 billion parameters per token, which means it punches at frontier weight while running at a fraction of the cost you would expect. It ships fully open under Apache 2.0, so anyone can pull the weights from Hugging Face, host them, fine-tune them, and put them straight into production.

The interesting part is not the size. It is where Hy3 lands on the benchmarks that teams actually care about, how far it jumped from its April preview, and what it signals about the direction of open-weight AI.

What Hy3 actually is

Hy3 uses a mixture-of-experts (MoE) architecture with 192 experts, of which the top 8 are activated per forward pass. The core network runs 64 layers with grouped-query attention, and a small 3.8B multi-token-prediction (MTP) layer sits on top to speed up generation. Total parameters: 295 billion. Active per token: 21 billion. Context window: 256K.

The design is a deliberate choice. Tencent’s team could have chased a trillion-parameter model, but past a certain scale multi-node deployment eats latency and throughput faster than raw capability grows. The 300B range keeps Hy3 servable on a single 8-GPU node with high-memory cards like the H20-3e, which matters if you want to run this yourself instead of renting it from a provider.

Deployment is straightforward. Hy3 works out of the box with vLLM and SGLang, exposes an OpenAI-compatible API, and can be tuned via LLaMA-Factory with DeepSpeed ZeRO. Reasoning depth is configurable per request through a reasoning_effort setting, so the same model can answer trivial queries quickly and burn extra tokens on hard math or agent workflows.

The jump from preview to release

Hy3 first appeared in April as a preview, and the gap between that snapshot and the full version is unusually wide. On SWE-bench Pro, which tests real software-engineering tasks, the score rose from 46.0 to 57.9. MathArena Apex climbed from 12.6 to 38.7. BrowseComp, the browsing-agent benchmark, moved from 67.1 all the way to 84.2.

A blind evaluation by 270 domain experts scored Hy3 at 2.67 out of 4, edging Zhipu’s GLM-5.1 at 2.51, which had been the reference point for open weights this year. Just as important, the internal hallucination rate dropped from 12.5% to 5.4%. That reduction is the sort of thing that decides whether a model gets deployed in a workflow or gets left as a demo.

Where it meets the closed frontier

The most interesting result is FrontierScience-Olympiad. Hy3 tops the board at 74.8, ahead of GLM-5.2 at 72.5 and GPT-5.5 at 73.8. On BrowseComp its 84.2 sits level with Claude Opus 4.8 at 84.3 and GPT-5.5 at 84.4. On agentic and scientific reasoning, in other words, an open-weight model has caught up with the closed labs.

The gaps that remain are honest ones. Claude Opus 4.8 still leads SWE-bench Pro at 69.2 against Hy3’s 57.9, and holds Terminal Bench 2.1 at 85.0. GPT-5.5 laps the field on MathArena Apex at 85.4 versus Hy3’s 38.7. So on the hardest coding tasks and the most brutal math benchmarks, the paid frontier is still ahead. Hy3 is not pitching itself as the best model on every axis. It is pitching frontier-adjacent capability you can host yourself.

How Tencent built it in 90 days

In February 2026, Tencent tore down its pre-training and reinforcement-learning infrastructure and rebuilt both from scratch. Training on the new framework started six weeks later. Ten weeks after that, Hy3 preview went live. The rebuild followed three principles:

  • Capability systematisation. No model should specialise its way out of product usefulness. Broad competence, not benchmark cherry-picking.
  • Evaluation authenticity. Test against real tasks pulled from live products, not just public leaderboards.
  • Cost-performance. Co-design the model and the inference framework so capability gains do not price it out of deployment.

The Hy model team merged with the Yuanbao chatbot team, CodeBuddy, WorkBuddy, ima, and QQ Browser into a single development loop. Live product metrics fed straight back into training priorities. That is how they landed things like emotional-tone calibration in the chatbot and multi-step scheduling reliability in the productivity assistant, both features that come from watching what users actually send rather than what a benchmark asks.

What it looks like in production

Inside Tencent’s own product stack, the early numbers are specific. In CodeBuddy and WorkBuddy, latency dropped 54%, end-to-end task duration dropped 47%, and success rates sit above 99.99%. Agent workflows of up to 495 sequential steps run stably in production. Those figures come from Tencent’s internal deployment, so treat them accordingly, but the direction is clear: this model is being pushed hard through real workloads.

Two examples show the reliability angle. Given meeting minutes with implicit start dates, leave arrangements, and overtime requirements scattered across multiple exchanges, Hy3 produced a correct, executable schedule without guessing at the missing pieces. On a multi-day travel plan, it juggled cross-day budgets, opening hours, and deduplication of stops without fabricating information. A model that refuses to invent an answer when context is incomplete is not doing anything glamorous, but it is doing the thing that determines whether teams trust it.

Pricing and access

API access through Tencent Cloud is priced at roughly one-tenth of GPT-4-class rates. The model is also available through OpenRouter, which routes requests across hosting providers based on whether you optimise for price, speed, or tool-calling accuracy. Because the weights are Apache 2.0, self-hosting is a real option for teams that want to avoid vendor lock-in or run sensitive data on their own infrastructure.

Why this release matters more than the individual scores

Every few months an open, self-hostable model gets closer to the paid frontier on the tasks most teams actually run. Hy3 is the clearest example so far of that pattern hitting the frontier itself, not just approaching it. FrontierScience-Olympiad leadership and BrowseComp parity with Claude Opus 4.8 and GPT-5.5 are not modest results. They mean that for a growing share of workloads, “good enough and open” is now a genuine alternative to a closed-model contract.

The story worth watching next is not the next leaderboard update. It is what this does to closed-model pricing. When frontier capability becomes forkable, the economics of the paid tier have to move. And with Tencent’s product loop feeding real interaction data back into training, the interesting question is whether the gap between open and closed keeps closing or whether the closed labs still have something structural the open ecosystem cannot copy.

The bigger shift underneath

One nuance that gets missed in the benchmark tables: Hy3 is not just an open model, it is an open model built inside a company that runs consumer-scale products. That feedback loop, product usage shaping training targets, is the harder thing to copy. Weights can be released. A live pipeline of millions of real queries feeding back into the next checkpoint cannot. Whoever wins the next round of open-weight models will be the team with the best data flywheel, not the biggest cluster.