Tencent Hunyuan Hy4 Preview: Apache 2.0 open weights and the school deployment reality

When a sponsoring body reviews open options, Hunyuan is often filed under "cleanest licence"; the part that gets underestimated is the 770B-parameter self-hosting bar — legal approval can pass while the hardware line item stalls for months.

Hunyuan puts both Hy3 and Hy4 Preview under plain Apache 2.0, which stands out among Chinese models on standard licensing. For Hong Kong schools the value is a clear legal path; the risks are Preview instability and the fact that Tencent Cloud API and self-hosting are two completely different compliance stories.

What Tencent is doing with Hunyuan

Hunyuan is Tencent's large-model family. Since Hunyuan-Large on 5 November 2024 — Tencent's first large open-weight MoE — the line has stayed on an open route. The 2026 Hy3 and Hy4 Preview releases use plain Apache 2.0 rather than bespoke vendor terms, which matters when GLM-5.3 flagship weights are still unpublished and Kimi K3 ships under a custom licence.

Training visibly targets workplace tasks: documents, spreadsheets, presentations, analysis. That maps onto the administrative tidying teachers actually face; Hong Kong curriculum detail, Traditional Chinese usage and DSE terminology still need local testing, not vendor blind benchmarks alone.

Models a school will actually see

Weights are free to download. Calling Tencent Cloud's API instead is billed at published rates, with input around $0.83 per 1M tokens:

An IT lead at a school group put it plainly: "Apache 2.0 cleared legal in two weeks; the GPU quote for Hy4 has been in front of the committee for three months." That tension is typical for Hunyuan at sponsoring-body scale.

Where the money comes from

Self-hosting: no licence fee for the weights; real cost is a server-class GPU cluster, power, staffing and model updates. Even though only 49B parameters activate per token, memory must hold all 770B — usually sensible only when shared across a sponsoring body, rarely for a single school server room.

Cloud: Tencent Cloud API bills per token at roughly $0.83 per 1M input tokens, trading hardware for a cross-border privacy assessment. Schools planning hardware with EDB AI funding (up to HK$500,000 per school, usable before 31 August 2028) may fit self-hosting into an "on-campus infrastructure" story; schools that want a quick trial may prefer cloud speed with heavier compliance paperwork.

Legal fees saved on Apache 2.0 may not cover the first year of a GPU cluster — evaluate hardware before the licence glow.

Its role inside Edor.ai

Not a native provider, but reachable through local Ollama if you self-host. 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.

Split responsibilities clearly: hardware, updates and faults on the self-hosted server belong to the school; platform integration, four-layer safety gating and ai_audit_logs belong to us. Document that split in the IT service catalogue before go-live — otherwise teachers assume Edor.ai fixes GPU outages, and the helpdesk inherits a problem it cannot touch. Embedding for the knowledge base and speech models for Cantonese oral practice still need an OpenAI or Azure OpenAI key or a separate local deployment. We can assess feasibility with you through contact us.

Scenarios that fit, and ones that do not

Better fit: a sponsoring body with or planning a GPU cluster that needs Apache 2.0 to pass the management committee; three-month pilots on non-critical admin tidying or meeting summaries; ICT work comparing MoE architecture with open deployment.

Picture a sponsoring body running Hy3 on shared GPUs for staff-meeting minute tidying — no pupil data, Apache 2.0 already approved, Edor.ai pointed at the Ollama endpoint for teacher tools. That is the shape of a sensible pilot. Moving the same setup to automated essay marking without re-testing rubric stability is not.

Poor fit: a single school with no server-class GPUs forcing Hy4 Preview into whole-school workflows. Wiring Preview straight into automated marking or outward-facing announcements. Assuming self-hosted Hunyuan covers embedding and speech — a chat model alone does not. Calling Tencent Cloud API on work containing pupil names without a cross-border assessment.

When the committee asks for a fallback, write it before the pilot starts: which OpenAI or Azure tier resumes if Hy4 latency spikes during exam season, and who owns the switch-over test. Preview releases are for learning, not for leaving the school without an exit ramp. Smaller schools should read Alibaba Qwen for single-machine trials while the sponsoring body negotiates cluster pricing — Apache 2.0 on Hy4 does not shrink the GPU bill.

The trade-off for schools

• If you have no GPU cluster roadmap yet, start with a smaller open model on Ollama rather than letting Apache 2.0 skip the hardware conversation.
• If data must stay in Hong Kong with a clear legal path, prioritise self-hosting and point Edor.ai at the Ollama endpoint instead of opening Tencent Cloud API first.
• If you pilot Preview, keep it on non-critical workflows, retain the ability to switch back to OpenAI or Azure, and test with your own rubrics rather than published benchmarks alone.