Tencent Hunyuan Hy4 Official Release Tested: Hallucination Rate Drops to 7.2%, Ranks First Among Open-Source Models in Financial Tasks
nashnova research
Tencent shipped the full release of Hunyuan Hy4 — 77 billion parameters, hallucination rate cut from 12.95% to 7.2% — and claimed the top open-source spot on financial tasks. But Tencent itself admits a sizeable gap with leading closed-source models on frontier math and long-horizon development.
How big is this model, and what can it do?
Hy4 full release carries 77 billion total parameters but activates only 4.9 billion per inference. This means → lower compute cost per query, which matters for high-volume deployment.
It adds native vision understanding and computer-use agent (CUA) support for the first time — the model can read images and operate desktop interfaces, not just process text.
This means → Hy4 has moved from a pure language model to a multimodal, action-capable agent, targeting automated workflows rather than chat alone.
How did the hallucination rate come down?
Hallucination rate dropped from 12.95% to 7.2%; speculative-error rate fell from 23.1% to 5.6% — both nearly halved.
In plain terms = "hallucination" is the model confidently fabricating facts; "speculative error" is the model guessing without evidence. Lower numbers mean more reliable output.
Tencent attributes the gains to significantly expanded pre-training and post-training data, plus targeted optimization for real-world productivity scenarios — not an architecture swap, but more and cleaner data.
Where do vision and agent capabilities stand?
On the vision side, Hy4 can locate ~2-meter rooftop components in a 2.8-billion-pixel aerial image — pixel-level perception.
In a hands-on test it parsed a design screenshot's colors, fonts, and layout, then generated an interactive front-end page. All 59 tests passed with zero console errors and zero runtime exceptions.
On long-horizon agent tasks, Hy4 can self-check mid-execution, locate issues, revise its plan, and re-verify until it delivers a bug-free result. This means → the model is developing a self-correction loop, not just one-shot output.
Why did it rank first among open-source models on finance?
Dedicated training covered fund research, securities analysis, quant strategy, asset allocation, corporate finance, and tax accounting — thousands of workspaces and tens of thousands of curated tasks.
The evaluation involved 213 domain experts, 70% from external top-tier financial institutions with five-plus years of experience — not an in-house self-grade.
Across 579 real financial tasks, Hy4 ranked first among open-source models. This reflects a heavy bet by Tencent on vertical financial data and human validation.
How does it compare with closed-source leaders?
Tencent claims Hy4 beat Claude Opus 5.5 on agentic coding, and beat both GPT 6.1 Sol and Claude Opus 5.5 on agentic search and work-agent tasks.
But Tencent itself concedes a sizeable gap with leading closed-source models on frontier math, long-horizon software development, model iteration, and visual Q&A.
In plain terms = on "good enough" work scenarios, Hy4 already competes with top closed-source models. On the hardest frontier benchmarks, the open-vs-closed gap is still wide open.
What comes next?
Hy4 was led by Yao Shunyu, a former OpenAI researcher who joined Tencent in late 2025 and reports directly to President Martin Lau — a reporting line that signals AI's strategic priority inside Tencent.
A lightweight Hy4 Flash is in development, targeting ultra-low inference cost for on-device deployment and high-concurrency agent orchestration.
The open-source weight release date has not been disclosed. This means → outside developers can only see benchmark scores for now; independent verification begins only when the weights ship — that is the real moment of truth.
市场有风险，内容仅供研究参考，不构成投资建议。