Tencent's open-weight line has turned over faster than most businesses review a software contract. Hy3 arrived in July as a free challenger to Claude. As at September 2026 its successor, Hy4, has already replaced it, and the more useful story is not the new model but what that pace does to anyone who built on the old one.
What is Tencent Hy4?
Tencent Hy4 is an open-weight large language model with 770 billion total parameters, of which roughly 49 billion are active for each token, released under the Apache 2.0 licence. It succeeds Hy3, which Tencent released in July 2026. Apache 2.0 means no royalty and no usage cap, the same permissive footing as Qwen's and Mistral's open lines, so a business can use it commercially without a separate agreement.
The split between total and active parameters is the mixture-of-experts design we unpacked when comparing Hy3 with Claude Opus 4.8. Only a slice of the model runs on any single request, which keeps inference cheaper than the headline size suggests. It does not make a 770-billion-parameter model easy to host. For almost every Australian mid-market business, using Hy4 means calling someone else's hosted endpoint.
The release cadence is the real headline
Hy3 to Hy4 in under two months is not unusual any more. Moonshot, Z.ai and DeepSeek are all shipping major point releases every four to eight weeks. For a vendor list, that creates the same three-step headache every time:
Your team picks a model version to fine-tune or build against
Eight weeks later a new version ships with different benchmark numbers and, sometimes, a changed licence
Documentation, community answers and hosting support drift toward the new version while you are still running the old one
The licence point is not hypothetical. We wrote earlier this year about MiniMax's mid-stream licence change: a model that was free for commercial use in June needed a different agreement by August. If your product logic is tied to Hy3, Hy4's release is the cue to read Tencent's terms again rather than assume nothing moved. Kimi K3 showed how much can sit in the fine print of an open-weight licence.
A quarterly review that fits on one page
The fix is not to avoid open-weight models. It is to treat each one as a fast-depreciating asset and put a review on the calendar. For a mid-market team we typically see each review cost $2,000 to $4,000 in engineering time, which is small next to the cost of discovering a licence change from a customer.
| Step | Question to answer | Typical owner |
|---|---|---|
| 1. Inventory | Which model versions are live, and where? | Engineering lead |
| 2. Licence check | Have the terms changed since we adopted this version? | Operations or legal |
| 3. Hosting check | Is our provider still serving this version, and from which region? | Engineering lead |
| 4. Re-test | Does the newer release beat ours on our own test set? | Engineering |
| 5. Decide | Stay, upgrade or move the workload to Claude? | Business owner |
Step four is where most of the cost sits, and it only works if you kept a test set from the first time round. If you did not, build one now: fifty real inputs with known good outputs is enough to catch a regression.
Where Hy4-class models belong in the stack
Our advice for Australian businesses has not changed with Hy4. Keep Claude as the default production layer wherever a client-facing answer needs to be right the first time. Treat Hy4-class models as candidates for internal, lower-stakes batch work, such as classifying support tickets, first-pass summaries of internal documents, or tagging archives, where a human reviews the output before anyone relies on it.
That split is the hybrid pattern more Australian businesses are running: a reliable managed model for the work that faces customers and regulators, and a cheaper open model for bulk work behind the scenes. Financial services firms answering to APRA or ASIC, in particular, should expect to explain which model produced which output and why.
The maths behind "free"
A 770-billion-parameter model with no licence fee sounds like a bargain until the engineering hours are counted. We have done that arithmetic for clients in Sydney and Brisbane this year. Once hosting, support and quarterly re-validation are included, the blended cost typically lands within 15 per cent of running Claude for the same workload. At that margin, reliability and a support path usually decide it.
There are exceptions. A business with very high, steady batch volume and an engineer who already maintains the pipeline can come out ahead. The ROI calculator is a reasonable first check, and our services page explains how we run the fuller comparison.
If you want a straight read on whether Hy4, or any other open-weight model, is worth chasing for your stack, book a call and bring the list of models you are currently running.
