Tencent expands Hy4 inference cluster tenfold after WorkBuddy usage surge

Tencent said on November 5 that it has expanded the inference cluster serving its Hy4 large language model tenfold, after a surge in usage on WorkBuddy, its AI agent for office work. Many users hit their quotas on the model's first day of availability.

Capacity and quota changes

According to a notice from the Tencent Hy and WorkBuddy teams, the company has also raised free usage quotas and will keep reallocating resources as demand shifts. The teams cautioned that limits may still appear at peak times because of constraints on high-end computing capacity and concentrated demand.

They suggested two workarounds: switching to other models such as Hy3, or moving non-urgent tasks to the evening. Free access to Hy3 on WorkBuddy has been extended to 23:59 on November 15.

Benchmark placement

Tencent cited the Intelligence Index that Artificial Analysis published on November 5, which ranks 724 models. The index combines evaluations spanning complex reasoning, coding, tool use, multimodal understanding, long-horizon planning and scientific problem solving.

• Hy4 placed second among Chinese models, behind Kimi K3.1 Pro and ahead of Qwen3.8 Max and DeepSeek V4.1 Flash.
• Hy4's score of 35 matches that of Grok 4.7 (xhigh).

On cost, Tencent said Artificial Analysis puts Hy4 at about $0.78 per task. The company describes this as roughly one-tenth the cost of Grok 4.7 (xhigh) and one-thirteenth that of Qwen3.8 Max, models in a similar score range. These ratios are Tencent's characterization.

Agent Arena

On Agent Arena, a leaderboard focused on real-world agent tasks, Hy4 ranks 13th overall. Tencent says this makes it the top open-source model on the board. The listing is currently marked preliminary.

Hy4 also sits on the leaderboard's Pareto frontier of score versus cost. Tencent says it keeps a positive score at a much lower cost per task than higher-priced models such as Claude Opus 5.5.

Hy4 background

Tencent released Hy4 on November 4. Compared with Hy4 preview, the model scales up pre-training and post-training data and adds visual understanding for the first time. Tencent also says it improves instruction following, long-context comprehension and long-horizon agent performance.

What this means

The capacity expansion shows that demand for Hy4 on WorkBuddy exceeded what Tencent had provisioned at launch. The teams themselves say limits may persist at peak times because of constraints on high-end computing capacity. Users running agent workloads on WorkBuddy should expect possible throttling and have Hy3, with its extended free access, as a fallback.

The benchmark claims centre on cost per task. Artificial Analysis puts Hy4 at about $0.78 per task, and Tencent frames that as a large cost advantage over models with similar Intelligence Index scores. The Agent Arena listing is still preliminary, so the 13th-place ranking and the top open-source claim may change as the board is finalized.