Tencent Hunyuan Hy4 Ships with Vision and Computer Use, But Third-Party Benchmarks Are Still Empty
For developers evaluating open-source models, Hy4's lack of third-party benchmark results and missing weights mean any production adoption depends on self-reported scores that may not transfer to real workloads, especially terminal-heavy tasks where it clearly lags rivals. The free window is a chance to test vision and long-context behavior, but don't expect it to replace a CLI-savvy agent.
Tencent released Hunyuan Hy4 as an official version on November 4th, keeping the same 770B-total/49B-active MoE architecture as the August preview. The real additions are perception and post-training: vision understanding for the first time, plus a Computer Use Agent (CUA) that can operate browsers, click buttons, and execute workflows across GUI and command line. The official claim of being "among the strongest open-source models" is defensible on Tencent's own twelve-item benchmark, where Hy4 scores eight third-place finishes, one second, and three fourths, with no first-place—but the numbers are self-reported, with four benchmarks labeled internal.
Cross-validation against third-party data is mixed. On Terminal Bench 4.0, official scores for GLM-5.3 and DeepSeek 1017 are within about 2 points of LLMBoard's November 3rd numbers, suggesting public test sets aren't mislabeled. But hallucination rates are reported inconsistently (15.76%→5.45% vs. 12.95%→7.2%), while the consistent figure is guessing-error rate dropping from 23.1% to 5.6%. The biggest weakness is terminal/CLI performance: Hy4 trails Kimi K3.1 Pro by more than half on Terminal Bench Science 0.1, and scores 44.4 on Terminal Bench 4.0 versus Kimi's 66.8.
Independent boards have not yet evaluated the official release; the preview ranked 14th domestically on one list, and 16th on Vals AI's Tax Agent Bench. Weights are not yet released—the preview was Apache 2.0, but the official version only promises "open-source soon." Pricing is unchanged at input ¥6/output ¥18 per million tokens, with limited-time free quotas on WorkBuddy and CodeBuddy through November 15th.
The official benchmark board is internally consistent but self-serving: no first-place finishes among five rivals strongly suggests Hy4 is a solid mid-tier model, not a leader, and the lack of any independent board slot keeps that conclusion provisional.
Tencent is positioning Hy4 for agentic workflows—vision and CUA are built for real-world GUI interaction—but the terminal results reveal a clear gap in command-line reasoning, which is where many RPA and ops tasks actually live.
Releasing weights late (or not at all, yet) while the preview was Apache 2.0 suggests Tencent may be hedging its open-source commitment even as it claims open-source leadership; the phrase 'recently open-source' without a date is a tell.
The free quota is intentionally small enough to prevent crowding but large enough to let serious testers hit limits quickly; the author's four-hour rate limit is a practical data point for API users.