Tencent Hunyuan (Hy4 Preview / Hy3): open weights under plain Apache 2.0, the easiest licence for a sponsoring body to approve

Short answer: Hunyuan has the cleanest licensing of the Chinese models covered here — Hy4 Preview and Hy3 both ship under plain Apache 2.0, which is the cheapest kind of licence to approve — but the self-hosting bar at 770B parameters is still there, and calling Tencent Cloud directly is a cross-border transfer.

What this is

Hunyuan is Tencent's large-model family. It has taken an open route since Hunyuan-Large in 2024, Tencent's first large open-weight MoE, and by the 2026 Hy3 and Hy4 Preview releases it had moved to plain Apache 2.0 rather than the bespoke licences common elsewhere.

That is worth a school's attention. Many recent "open" models in fact carry territorial limits, revenue thresholds or extra commercial conditions that a legal adviser must read line by line. Apache 2.0 is an international standard with clear wording and an explicit patent grant, so the review work is much smaller. With the GLM-5.3 flagship weights still unpublished and Kimi K3 under a bespoke licence, Hunyuan is genuinely ahead on this point.

The current line-up

• Hunyuan Hy4 Preview (weights published 28 August 2026): 770B total parameters with 49B active per token, a 1M-token context window, under Apache 2.0. Tencent's own blind evaluation places it slightly ahead of GLM-5.3 and Kimi K3. The Preview label is Tencent's own, and the model card states plainly that this is an early version with known issues.
• Hunyuan Hy3 (6 July 2026): the full release, also Apache 2.0 and distinct from the more restrictive April Hy3-Preview. In late July 2026 it ranked third by token volume on OpenRouter, making it one of the few open-weight models with demonstrable real-world usage at scale.
• Hunyuan-Large (5 November 2024): Tencent's first large open-weight MoE and the start of Hunyuan's open-source direction.

The weights themselves are free to download; calling the Tencent Cloud API instead is billed at its published rates, with input around $0.83 per 1M tokens. The full version comparison is in the automatically generated series table further down this page.

Strengths

1. The easiest licence to approve. Apache 2.0 is an international standard with clear commercial, modification and redistribution terms plus a patent grant. For anyone who has to explain legal risk to a management committee, that is the most practical advantage on this page.
2. Capability in the frontier band. 770B total parameters with a 1M-token context window puts it at the top of the open-weight field, and while the blind evaluation is Tencent's own, it points the same way as independent usage data.
3. Evidence of real usage. Hy3 ranking third by token volume on OpenRouter in late July 2026 shows it is not just a launch-post model but one people run continuously.
4. Trained around real work. The training data is visibly built around practical tasks — documents, spreadsheets, presentations, analysis — which maps directly onto the administrative and data-tidying work teachers actually face.

Limits

• Self-hosting still needs server-class hardware. Only 49B parameters activate, but memory has to hold 770B, so the real requirement is a GPU cluster rather than a school workstation. Licence freedom is not deployment feasibility.
• The Tencent Cloud API is a cross-border transfer. Without self-hosting you are calling the cloud, data leaves Hong Kong, and the school owes its own privacy assessment and parent notification arrangements.
• The Preview release is still moving. The known issues include over-long reasoning on complex tasks and repeated self-verification, so output length and latency are not fully predictable — a reason not to migrate critical school workflows in one step.
• No embedding or speech companion. Even with a self-hosted chat model, the vector index behind the knowledge base and Cantonese oral practice need separate solutions.
• A thin education ecosystem. No Hong Kong curriculum prompt templates or teacher community, so an IT coordinator has to build internal documentation and training material from scratch.

How a Hong Kong school should view it

Hunyuan deserves a serious look from sponsoring bodies, and the reason is the licence: it removes any ambiguity from the sentence "we may lawfully run this model on our own servers". When evaluating open options, that saves more than legal fees; it saves rounds of clarification.

The evaluation order should not be reversed, though. Ask about hardware first: is there a GPU cluster, or a plan for one? Who maintains it? What is the fallback during downtime? If those three have no answer, the licence does not help and the discussion should move straight to a smaller open model. Ask about compliance second: if the decision lands on cloud, write down the cross-border assessment, retention period and access rights; if it lands on self-hosting, that gate largely disappears, which is precisely the value of self-hosting. Only then ask about capability, and test it with the school's own lesson plans, worksheets and marking criteria rather than published benchmarks.

One closing caution: a Preview release suits a pilot, not a wholesale migration. A sensible approach is three months on one or two non-critical workflows — administrative document tidying, say — while keeping the ability to switch back to the original provider at any time.

Availability inside Edor.ai

Not a native provider, but reachable through local Ollama if you self-host. Edor.ai natively supports OpenAI, Azure OpenAI, Anthropic, Poe and local Ollama. We do not build in a Tencent Cloud connector and we have no partnership with Tencent. If a school or sponsoring body stands up an inference server and exposes an OpenAI-compatible endpoint through Ollama, the platform can point at it and teacher tools, the AI tutor and student-conversation gating all continue to work.

Responsibilities split cleanly: hardware, model updates and fault handling on the self-hosted server belong to the school, while platform integration, the four-layer safety gating and the ai_audit_logs audit trail belong to us. Embedding for the knowledge base and speech models for oral practice still need an OpenAI or Azure OpenAI key or a separate local deployment. We are happy to assess feasibility with you through contact us.

Alternatives

• Limited hardware budget and a wish to self-host on a single machine — see the 27B class under Alibaba Qwen.
• Equally standard open licensing with shorter terms — see the MIT releases under Zhipu GLM.
• To work out the cost gap between self-hosting and cloud, read the open versus closed cost and privacy trade-off and see how the EDB AI funding fits.
• For a provider that works immediately, start with OpenAI or Anthropic Claude, or go back to the model overview.