Tencent Hy4: Open-Weight Model Gets Vision and Computer Use
Tencent’s Hunyuan team released Hy4 in general availability on November 4, 2026, delivering the most capable open-weight model the lab has ever shipped. The release is significant for enterprise AI teams for two reasons: it adds vision and computer use to a model that was already competitive on agentic benchmarks, and it promises Apache 2.0 weights that will make GPU-cluster self-hosting straightforward.
What launched
Hy4 is a Mixture-of-Experts model with 770 billion total parameters and 49 billion active, running over a 1 million token context window. It is now available via Tencent Cloud and OpenRouter at $0.834 per million input tokens and $2.501 per million output tokens, with cached input at $0.042. The API is OpenAI-compatible, meaning most teams can point their existing client at a new base URL with no other changes.
The GA build improves substantially over the August 2026 preview, which introduced the model but lacked vision and was not available as open weights. The new additions are:
- Vision understanding. Hy4 now processes images, analyzes complex charts, and performs visual mathematical reasoning.
- Computer use. The model operates command-line interfaces and graphical UIs for the first time, with targeted training for professional software including Blender and CAD as well as multi-application workflows.
- Hallucination reduction. The hallucination rate dropped from 15.76% to 5.45% in Tencent’s internal evaluation, a material improvement for production deployments where reliability is non-negotiable.
Enterprise benchmark performance
Tencent ran a 579-task real-world finance evaluation with 213 finance professionals, 70% of whom had over five years of industry experience. Hy4 scored 3.04 and ranked first among open-weight models, ahead of DeepSeek V4.1 Flash (2.97), Kimi K3.1 Pro (2.68), and GLM 5.3 (2.68).
The table below summarizes open-weight model performance across the finance evaluation:
| Model | Finance Score | Hallucination Rate |
|---|---|---|
| Tencent Hy4 | 3.04 (1st) | 5.45% |
| DeepSeek V4.1 Flash | 2.97 | n/a |
| Kimi K3.1 Pro | 2.68 | n/a |
| GLM 5.3 | 2.68 | n/a |
For coding work, GameDev improved 19.1% and GameLogic improved 9.8% over the preview. On the internal multi-turn evaluation suite, the overall score rose from 65.25 to 72.53, with multi-turn memory up 6.67 points and writing quality up 10.89 points.
Tencent is transparent that a gap remains versus top proprietary models on research-level mathematics, long-horizon coding, model self-improvement, and advanced visual QA. This is consistent with where other open-weight labs have positioned their releases: genuinely useful for a wide range of enterprise tasks, still behind the absolute frontier on the hardest benchmarks.
Five domains, one model
The Hunyuan team designed Hy4 around five distinct use cases rather than chasing a single headline benchmark.
Vision and computer use. For teams running computer use agents, Hy4 supports both CLI and GUI operation. The model was trained on long-horizon workflows that span multiple applications, so agents can plan multi-step tasks without losing context. This matters for enterprise automation use cases such as back-office data entry, desktop application testing, and document processing pipelines that require reading a rendered interface.
Frontend and game development. Visual understanding drove a step change in generated UI quality: Tencent says Hy4 produces interfaces that are functionally complete and production-ready, not just visually polished. For game development, Tencent partnered with professional developers inside its Interactive Entertainment Group to curate training data and evaluations for Unreal Engine, Unity, and Godot, including multiplayer game development.
Finance and office workflows. Tencent built thousands of realistic workspaces and tens of thousands of tasks covering fund research, equity analysis, quantitative strategies, asset allocation, and corporate finance. WorkBuddy integrates with WeCom to surface chat history, emails, documents, and calendars so the model has the context it needs to manage multiple projects. Hy4 also supports co-editing Word, Excel, and PowerPoint files in real time, with improvements in document formatting and reliability over the preview.
Scientific research and AI R&D. On TenureTrail, Tencent’s internal benchmark of 123 research tasks across seven disciplines, Hy4 ranks first among open-weight models. The research angle extends to recursive self-improvement: paired with the Hyra 2.0 harness, Hy4 optimized Tencent’s internal pre-training infrastructure for smaller models and autonomously trained a variant that outperforms the official Jev baseline on multiple benchmarks. Tencent is open-sourcing two foundation models co-developed through this loop: HyCell for single-cell representation and HyPDE for physical field modeling.
Long-horizon personal assistance. The WorkBuddy integration connects Hy4 directly to employees’ work context including calendars and meeting notes, enabling multi-project tracking, action-item surfacing, meeting briefs, and deadline prioritization. Team collaboration features built on Hy4 will roll out in WorkBuddy alongside the new Projects feature.
Why open weights matter for enterprise AI teams
The commitment to open-weight release under Apache 2.0 is the most consequential element for enterprise buyers running their own AI infrastructure. Apache 2.0 permits commercial use, fine-tuning, and redistribution without royalties, making Hy4 a candidate for teams that need to run models in air-gapped environments, customize on proprietary data, or avoid per-token API pricing at large scale.
The reference self-hosting configuration uses 16x B200 GPUs. Tencent is partnering with the vLLM and SGLang communities to provide deployment recipes, which means most infrastructure teams will be able to get the model running with tooling they already use. A system card with safety evaluations will accompany the weights.
This follows the pattern set by other major open-weight releases in 2026: Mistral Large 4, Kimi K3.1 Pro, and DeepSeek V4.1 Flash all shipped with permissive licenses and vLLM/SGLang support. The convergence is making it substantially easier for enterprise teams to evaluate and run frontier-class open-weight models in production without specialized MLOps expertise.
Comparison with the competitive field
The November 4 release lands in a crowded field. Kimi K3.1 Pro’s open weights dropped on the same day, also carrying a 2.8T-total parameter MoE count at competitive pricing. Mistral Large 4 (Le Chonk) released open weights in late October with a 1T-parameter multimodal MoE at $1.36/$4.18 per million tokens. DeepSeek V4.1 Flash 1017 offers strong agentic benchmark numbers at $0.15/$0.60 per million tokens.
For enterprise buyers evaluating these models, the practical differentiation points are:
| Factor | Hy4 | Kimi K3.1 Pro | DeepSeek V4.1 Flash |
|---|---|---|---|
| Total params | 770B | 2.8T | Undisclosed |
| Vision | Yes | Yes | No |
| Computer use | Yes | No | Limited |
| API price (input/output) | $0.834/$2.501 | $2.00/$10.00 | $0.15/$0.60 |
| Open-weight license | Apache 2.0 (pending) | TBD | MIT |
| Finance benchmark rank | 1st (open-weight) | 2nd | 3rd |
Hy4’s pricing positions it in the mid-tier: more expensive than DeepSeek but substantially cheaper than Kimi K3.1 Pro. Its lead on the finance benchmark and its computer use capability give it a differentiated story for enterprise deployments that need desktop automation alongside language tasks.
What to watch
Tencent flagged two near-term releases. Hy4 Flash is coming soon with significantly improved cost-efficiency and faster inference, designed for high-volume production use cases where the full 770B model is cost-prohibitive. The open-source weights release will follow the GA launch, bringing the first full public audit of the model’s training and safety posture.
For enterprise AI teams deciding whether to adopt Hy4, the checkpoint to watch is the system card and license confirmation. Apache 2.0 without restrictions is what most legal teams want to see before approving model adoption. The vLLM and SGLang support will make the infrastructure integration straightforward once weights drop.
Teams building multimodal workflows that need both vision understanding and desktop automation now have a credible open-weight option from Tencent’s enterprise product ecosystem. That is a narrower set of requirements than a pure-language task, but for the right deployment it represents a meaningful capability step over the preview model that launched three months ago.
If your team is evaluating open-weight models for enterprise deployment and needs guidance on selecting the right model for your stack, talk to Enera.