The latest “revolutionary” AI release mostly revolutionizes the press release cycle. But every once in a while, a model drops that actually changes the calculus for developers who build on top of these systems. Tencent’s Hy4, announced November 4, 2026, is one of those moments, even if the benchmark numbers deserve a skeptical eye.
Here’s the short version: Tencent just shipped a 770-billion-parameter Mixture-of-Experts model with a million-token context window, native vision understanding, and computer-use capabilities. The weights are Apache 2.0 licensed. And the company claims its model line-up now ranks among the strongest open-weight systems available today.
That last claim? Take it with a grain of salt until independent benchmarks land. But the strategic signal is unmistakable: China’s tech giants are all-in on open-weight AI, and Tencent just bought a seat at the table.
The Spec Sheet That Actually Matters
Before we get into the strategic implications, let’s look at what Hy4 actually brings to the table. The preview version, which shipped in late August, is a Mixture-of-Experts model with these specs:
| Spec | Hy4 Preview |
|---|---|
| Total parameters | 770 billion |
| Active parameters per token | 49 billion |
| Context window | 1,048,576 tokens |
| Layers | 78 (1 dense, 77 MoE) |
| Experts per MoE layer | 256 routed + 1 shared |
| Experts active per token | 8 routed + shared |
| License | Apache 2.0 |
The math here is worth unpacking. With 770 billion total parameters but only 49 billion active per token, roughly 6.4% of the network does work on any given inference. That’s the elegance of MoE architecture, you get the knowledge capacity of a massive model with the inference cost of something much smaller.
But here’s where the hardware reality bites. The model doesn’t care about active parameters when it comes to memory. You need to fit all 770 billion weights in memory, which means:
- BF16/FP16: ~1.54 TB
- FP8 (official build): ~770 GB
- 4-bit quantized: ~385 GB (community builds, no official release)
Those numbers don’t include the KV cache, which gets nasty with a million-token context. If you’re planning to self-host, you’re looking at a serious multi-GPU node. This isn’t a laptop toy.
Vision and Computer Use: The First-For-Tencent Features
The full Hy4 release adds two capabilities that the preview didn’t have. Vision understanding is the headline, the model can process images, analyze complex charts, and handle visual mathematical reasoning. Combined with the million-token context window, it can read a financial report with embedded graphs, or analyze a dashboard screenshot in one pass.
More interesting is computer use. Tencent trained Hy4 on professional software like Blender and CAD, and the model can interact with both CLI and GUI interfaces. The company claims it can plan and execute long-horizon tasks across multiple applications. For agent developers, this is a big deal, an open-weight model that can navigate a desktop environment is rare.
The game development angle is particularly Tencent. Working with developers from its Interactive Entertainment Group, the company tuned Hy4 for major engines like Unreal Engine, Unity, and Godot. The reported gains of 19.1% on GameDev and 9.8% on GameLogic over the preview are impressive on paper, but these are internal benchmarks. The baseline numbers haven’t been independently validated.
The Financial Workflow Play: Numbers That Sound Good, Need Scrutiny
Tencent’s most detailed performance claims come from the financial sector, where it ran a blind evaluation with 213 finance professionals across 579 real-world tasks. The results:
| Model | Score |
|---|---|
| Hy4 | 3.04 |
| DeepSeek-V4.1-Flash-1017 | 2.97 |
| Kimi K3.1 Pro | 2.68 |
| GLM-5.3 | 2.68 |
Hy4 edges out DeepSeek’s offering but doesn’t blow it away. More interesting is the operational improvement story. Tencent says targeted search optimization cut the hallucination rate from 15.76% (preview) to 5.45%, and reduced speculative errors from 23.1% to 5.6%. Those are massive improvements for retrieval-augmented workloads, if they hold up in independent testing.
There’s a pattern here worth noting. Tencent isn’t just building a model, it’s building a model-product loop. The Hy4 release is co-designed with WorkBuddy, WeCom, and Tencent’s cloud infrastructure. The internal evaluations show an average score of 83.18 and 96.15% pass rate on WorkBuddy tasks, ranking first in 9 of 12 scenarios in mainland China. When “co-design” means the model is optimized to sell Tencent’s enterprise ecosystem, the benchmarks measure what Tencent wants to sell, not necessarily what you want to build.
The Open-Weight Arms Race: Why Tencent Had to Play
Here’s the strategic context that makes Hy4 more than just another model release. Tencent is following Alibaba into the open-weight arena, and the timing is deliberate. The open-weight AI race with DeepSeek, Z.ai, and Alibaba has been heating up all year, and Tencent’s 770B parameter count represents a 2.6x jump from the 295 billion in Hy3.
Why does open-weight matter so much? It comes down to cloud economics. When a developer downloads a model’s weights, they still need somewhere to run it. Tencent Cloud TokenHub becomes the natural home for enterprises that want managed inference without the hassle of self-hosting a 770B parameter monster. Open weights are a funnel, the cloud is the revenue.
The pricing strategy confirms this. Hy4 API pricing sits at $0.834 per million input tokens and $2.501 per million output tokens, unchanged from the preview despite significant performance gains. Tencent Cloud is offering 50% off through December 31, 2026. That’s an aggressive land-grab strategy, undercutting closed proprietary APIs while competing on capability with other open-weight providers.
The company’s capex numbers tell the real story. Tencent’s capital expenditure jumped 176% year-on-year to 52.8 billion yuan ($7.9 billion) in Q2 to fund AI efforts, pushing free cash flow negative. This is a company betting its balance sheet on AI infrastructure. The open-weight strategy is how it recoups that investment.
The Cybersecurity Culling: Who’s Really Getting Neutered?
Now for the uncomfortable part. Tencent explicitly filtered Hy4’s training data to limit offensive cybersecurity capabilities. The official release notes state:
During Hy4’s training, we limited cybersecurity data to defensive material related to proof-of-concept (PoC) generation and vulnerability patching. We filtered out offensive content such as exploit development, reverse engineering, and penetration testing to avoid amplifying potential offensive capabilities.
This follows the same pattern as Kimi K3.1 and other recent Chinese releases. The stated rationale is safety, preventing the model from being a weapon-building toolkit. The actual effect is more complicated.
There’s a real tension here that developers should recognize. The same filtering that prevents Hy4 from helping with exploit development also strips capabilities that legitimate security researchers use daily. Defensive security work often requires understanding offensive techniques. A model that can’t reason about reverse engineering is a model that’s less useful for identifying vulnerabilities in the first place.
The developer community has noticed. The sentiment on technical forums is that Zhipu has been doing this for months, and DeepSeek has started too. The pattern is clear: Chinese open-weight labs are competitively culling cybersecurity capabilities to avoid regulatory scrutiny while still shipping useful models.
The uncomfortable implication is that security researchers now face a choice. Run open models with reduced security capabilities, or pay premium prices for closed models that maintain full offensive reasoning. The market for “non-gimped cybersecurity models” is already developing, and it’s expensive.
The Blunt Truth About Benchmarks
Here’s what nobody in the press release wants to emphasize: Tencent itself admits the gap to top proprietary models remains.
The official Hy4 research post acknowledges “a notable gap remains between it and top proprietary models, particularly on challenging internal benchmarks for research-level mathematics, long-horizon coding, model self-improvement, and advanced visual question answering.”
The Terminal Bench numbers are illustrative. Tencent reports Hy4 scores 44.4 on Terminal Bench 4.0, compared to 66.4 for Claude Opus 5.5 and 55.5 for GPT-6.1 Sol. On OSWorld 2.1 (computer use), Hy4’s 78.9 sits just behind Claude’s 81.8 and ahead of GPT’s 75.2. So it’s competitive in some agentic tasks but significantly behind on others.
Independent verification is completely absent. The benchmarking service Artificial Analysis has not published measurements for Hy4, and the model doesn’t appear in its open-model ranking. That ranking is currently led by DeepSeek V4.1 Flash (33 points) and GLM-5.3 (32 points), with open Qwen3.8 at 27B scoring 20.
The gap between vendor-reported and independent benchmarks can be stark. Mistral Large 4 reported 41% on Terminal-Bench, while Artificial Analysis measured 25%. Until someone independently evaluates Hy4, treat all the benchmark claims as marketing material with technical flavor.
The Hardware Reality Check
If you’re considering self-hosting Hy4, the hardware math should give you pause. As we covered in our analysis of open-weight AI models, the infrastructure requirements for frontier-scale MoE models are substantial.
For Hy4 specifically:
| Precision | Bytes/Parameter | Approx. Weight Size | Practical Deployment |
|---|---|---|---|
| BF16/FP16 | 2 | ~1.54 TB | Multi-node cluster |
| FP8 | 1 | ~770 GB | 8x H100 or equivalent |
| 4-bit | 0.5 | ~385 GB | 4x H100, tight fit |
These estimates exclude the KV cache, which grows with context length. A million-token context window is a feature you’ll rarely use in practice because the memory cost is astronomical. Most deployments will use 32K-128K context windows and accept the limitation.
The practical advice: unless you have serious GPU infrastructure, use the hosted APIs. The 50% discount through December 31, 2026 makes the exploration cost manageable. Self-hosting should wait until you’ve validated the model on your workloads and quantified the infrastructure cost.
What Developers Should Do This Week
The Hy4 release has real utility for teams building AI applications, but not for everyone. Here’s a practical path forward:
- 1. Test through OpenRouter. If you already route traffic through OpenRouter, swapping in Hy4 takes minutes. Run 50-100 real prompts from your workload and compare against your current model. Focus on cost, latency, and output quality, not benchmark leaderboards.
- 2. Try the free WorkBuddy/CodeBuddy window. From November 4-15, these tools are free. This is the lowest-friction way to evaluate Hy4 in production-like workflows, especially for code generation and document tasks.
- 3. Wait for the system card. The full weights release will include a comprehensive safety evaluation document. Before self-hosting, review this to understand what capabilities were filtered and whether those gaps matter for your use case.
- 4. Watch for independent benchmarks. Within weeks, third-party evaluations will land. Those will tell you whether Hy4’s claims hold up. Don’t commit infrastructure spending based on vendor-reported numbers.
- 5. Consider Hy4 Flash when it ships. The lighter-weight version promises better cost-efficiency with the same vision and agentic capabilities. For production workloads at scale, this will likely be the more practical option.
The Bigger Picture: Open-Weight AI Is Getting Crowded
Hy4’s release marks a pivotal moment in the open-weight AI landscape. With Tencent, Alibaba, DeepSeek, Z.ai, and Moonshot all shipping frontier-scale open models, the competitive pressure on proprietary APIs is intensifying. The open-weight AI hitting 95% of frontier performance at 5% cost dynamic is real, and it’s accelerating.
For enterprises, the implications are significant. Open-weight models provide data sovereignty, customization flexibility, and cost predictability. The trade-off is infrastructure complexity and security responsibility. That’s a different calculus for every organization.
For Tencent specifically, Hy4 is a statement of intent. The company is no longer content to be a consumer of open-source AI, it’s becoming a major producer. The model-product co-design strategy, combined with aggressive cloud pricing, positions Tencent to capture a significant share of the enterprise AI market in China and potentially beyond.
The honest assessment is that Hy4 is a strong open-weight contender that sits at the top of the pack in specific categories, financial workflows, game development, and computer use, while trailing proprietary leaders in research mathematics and long-horizon coding. Whether that’s enough to disrupt your current AI stack depends entirely on your workloads.
One thing is certain: the open-weight map just got more crowded, and that’s good news for developers who value choice.
