Tencent Hy3 is the third generation of the Hunyuan model family and has been available as a full release since July 6, 2026 – for the first time under a genuine Apache 2.0 licence without territorial restrictions. That is the real news for European companies: the April preview’s licence still explicitly excluded use in the EU, the UK and South Korea; the final release lifts this restriction. With Hy4-Preview (28 August 2026), a considerably larger open-weights preview (770B vs. 295B total parameters) is also available under Apache 2.0; Hy3 remains the stable reference for production self-hosting. This overview reflects the state as of October 2026.
Architecture: Efficiency over Size
Hy3 is a mixture-of-experts model with 295 billion total and only 21 billion active parameters (192 experts, 8 active per token). Add a 256K token context window, an additional multi-token prediction layer for faster decoding, and a hybrid fast/slow thinking design that toggles between quick replies and deep reasoning. Besides the standard version, an FP8-quantised variant is available for efficient self-hosting.
The positioning is clear: Hy3 is the efficiency play among Chinese open-source models – considerably smaller than DeepSeek V4-Pro (1.6 trillion parameters) and GLM-5.2 (744 billion), with competitive results in most disciplines except coding.
Benchmarks
According to Tencent and independent evaluations, Hy3 achieves, among others, 90.4 on GPQA Diamond, 78% on SWE-bench Verified, 57.9% on SWE-bench Pro and 84.2 on BrowseComp. On demanding maths (MathArena Apex: 38.7 vs. 85.4 for GPT-5.5) and in coding, the model trails the Western flagships and GLM-5.2. For the model size – only 21B active parameters – the results are remarkable.
Hy4-Preview: the Next Generation
With Hy4-Preview, Tencent released the successor as an open-weights preview on August 28, 2026 (Hugging Face: tencent/Hy4-preview, plus an FP8 variant tencent/Hy4-preview-FP8). Both are listed under Apache 2.0 per the Hugging Face model card – unlike Hy3, there is no territorial restriction from the outset.
Key specs (vendor-reported):
- 770B total, 49B active parameters (MoE) – considerably larger than Hy3
- Context window over 1M tokens
- In an internal blind evaluation by Tencent (163 experts, 203 engineering tasks), Hy4-Preview scores 2.99 out of 4.00, narrowly ahead of GLM-5.3 (2.92) and Kimi K3 (2.94)
- Integrated into Tencent’s own products WorkBuddy, CodeBuddy, Yuanbao and ima from launch
- API via Tencent Cloud TokenHub: per the vendor, around $0.834/1M input, $2.50/1M output, $0.042/1M cache-hit
Context: this is explicitly a preview, not a final full release. The comparison figures against GLM-5.3 and Kimi K3 come from an internal Tencent evaluation; we have no independent third-party confirmation. For production self-hosting deployments we therefore continue to recommend Hy3 as the established, stable full release, while watching how Hy4 develops.
Availability and Cost
The weights are available on Hugging Face and ModelScope; as an API, Hy3 is offered via Tencent Cloud’s TokenHub (~$0.18 input / $0.59 output per 1M tokens) and via OpenRouter. Within Tencent’s own ecosystem, Hy3 powers the WeChat assistant, Yuanbao and CodeBuddy, among others.
GDPR and Data Sovereignty: Self-Hosting Yes, China API with Caution
For European companies, the same logic applies as with DeepSeek and Qwen: Hy3 is interesting above all as an open-weights model for self-operation. Under Apache 2.0 it can be hosted in your own cloud environment – on STACKIT, Azure or AWS, for example – meaning no data flows to the vendor and full data sovereignty is maintained. The API via Tencent Cloud, by contrast, means processing by a Chinese provider and should be assessed critically from a GDPR perspective for personal or confidential data.
When reviewing the licence, mind the version: only the full release of July 6, 2026 and Hy4-Preview from August 28, 2026 are under Apache 2.0 – anyone still using artifacts of the April 2026 Hy3 preview operates under the old community licence with its EU exclusion.
Hyperscaler catalogues: We found no native offering for Hunyuan on AWS Bedrock (no model entry in the Bedrock model cards), Google Vertex AI (the corresponding Model Garden documentation page returns a 404 error), or Azure AI Foundry. Unlike GLM, Kimi or MiniMax, Hunyuan is therefore visible neither directly nor via Fireworks partnerships in the major hyperscaler catalogues; self-hosting remains the only route to EU data processing that we are aware of.
Positioning Among Chinese Models
| Model | Size (active) | Licence | Distinctive feature |
|---|---|---|---|
| Tencent Hy4-Preview | 770B (49B) | Apache 2.0 | Newest model, still preview status |
| Tencent Hy3 | 295B (21B) | Apache 2.0 | Efficiency, most permissive licence |
| DeepSeek V4 | 1.6T (Pro) | MIT | Reasoning leader among open models |
| GLM-5.2 | 744B | MIT | Coding strength |
| Qwen 3.7 | API-only (Max) | proprietary (Max) | Alibaba ecosystem |
Integration with CompanyGPT
As an open-weights model, Hy3 – like Llama, Qwen or GptOSS – can be integrated into a self-hosted platform such as CompanyGPT and operated GDPR-compliantly in your own cloud. For most enterprise use cases in the DACH region, we continue to recommend the established frontier models via EU endpoints; Hy3 is an exciting option for cost-efficient, sovereignly hosted workloads.
For individual consulting on the right model strategy, contact innFactory AI Consulting.