Hy4 Ranks 14th: Open-Source No. 1 Is a Futures Contract; $0.55 per Task Is Spot

Tencent Hunyuan released Hy4 on November 4. Arena.ai's Agent Arena ranks it 14th overall. This leaderboard is based on long-horizon agent sessions from more than 8.5K real users, which is not the same as single-question lab tests. In these sessions, Hy4 had a net improvement of 4.7% and entered Agent Arena's Pareto frontier. Arena also gives a more practical coordinate: $0.55 per task, versus $1.46 per task for Kimi K3.1 Pro Max, about 62% cheaper, with the score difference between the two within 1 percentage point. The leaderboard and cost page are at Agent Arena.

Beyond ranking, what Hy4 fills in this time are shortcomings in agent scenarios. Arena lists four signals: user-confirmed success ranks 11th, +8.7%, preview version 16th, +4.9%; positive-to-complaint ratio ranks 18th, +3.8%, preview version 24th, +1%; steerability in response to user corrections ranks 14th, +5.7%, preview version 27th, +0.21%; Bash reliability ranks 20th, +6.2%, preview version 41st, -3.1%. Compared with the preview version's 25th place and +1.1%, this is an overall improvement of 3.6 percentage points. One commenter replied that moving Bash reliability from 41st to 20th is the real jump; a 62% price gap plus a 1-point score gap is a number agent routing will take seriously.

Release information comes from the Tencent Hy account: Hy4 adds visual capabilities, covering finance, front-end, games, office, and research, and emphasizes being built according to cybersecurity and biosecurity requirements. Link at Tencent Hy4 introduction. AGTP's summary adds parameters: Hy4 is a 770B-total-parameter MoE, activating 49B per token, 2.6 times the previous generation Hy3 295B; 1M token context, 64K output; Tencent's own test APEX-Agents 75.2, DeepSeek V4.1 Flash between 72.1 and 72.9, with mixed results on other benchmarks. The final Hy4 is already available on Tencent Cloud, OpenRouter, and WorkBuddy, but open-source weights have not been released; Tencent says coming soon, and a smaller Hy4 Flash will follow. The Hy4 preview version was released with Apache 2.0 open-source weights on August 28, so this time the community is watching the words 'open source' very closely.

The controversy is here too. Arena writes that Hy4 is currently still a proprietary model; once weights are released, it will tie with Kimi K3.1 Pro for No. 1 among open-source models. Some netizens bluntly say that until weights land, it cannot be called the open-source crown; others offered a more cutting phrase, calling it a Schrödinger's open-source model. Tencent Hy only left one line in reply: Will Open Sooooon. On the other side, some netizens questioned that Arena's bar chart only shows increments and does not mark the baseline, making it easy for readers to mistake incremental differences for capability differences. This reminder is useful when reading leaderboards; when judging, still go back to scores, costs, and task types.

For people using agents for work, what Hy4 is more worth looking at right now is $0.55 per task, improved Bash reliability, and the ability to respond to user corrections. 14th place is only a leaderboard coordinate. Before open-source weights are released, 'open-source No. 1' is a futures contract to be delivered; the API is already usable, and that is spot. If you want to try it, Tencent Cloud, OpenRouter, and WorkBuddy all work. Cited sources include Arena.ai tweet, Tencent Hy tweet, and AGTP summary.