Tencent launched a new model today, and it is not a bigger flagship. It is the opposite: Hy4 flash, a smaller, cheaper sibling of the Hy4 flagship Tencent released less than two weeks ago. The pitch is simple — keep the 1-million-token context and the multimodal inputs, but cut the per-task cost to one-twentieth of the flagship's.
If that sounds too good to be true, the details matter. Here is what Tencent actually announced, what the model is, and where the trade-offs are.
What Hy4 flash is
Hy4 flash is a 224-billion-parameter model built on Tencent's Hunyuan architecture. Only 12 billion parameters activate per forward pass. Of the 224 billion total, 64 billion are Engram parameters — a conditional memory technique that stores frequently occurring text patterns in a fast-lookup table, so the model retrieves them instead of recomputing them.
For comparison, the full Hy4 flagship carries 770 billion total parameters with 49 billion activated per token. Hy4 flash is roughly a third of the flagship's total size and about a quarter of its activated parameters.
What it keeps:
- Multimodal inputs — text and images in, text out
- A 1-million-token context window — the same long context as the flagship
- Agent-oriented design — the Hy series is built for multi-step workflows, document work, and long-horizon tasks
The model is aimed at everyday work rather than frontier benchmarks: research, slide creation, spreadsheet editing, and the kind of long document and code-adjacent tasks that fill a workday.
The price is the story
Hy4 flash's list API prices are RMB 0.5 per million input tokens, RMB 2 per million output tokens, and RMB 0.02 per million cached input tokens. In dollar terms, that is roughly $0.07 per million input tokens, $0.281 per million output tokens, and $0.003 per million cached tokens.
Against Hy4's list prices, those rates represent reductions of about 12x on input, 9x on output, and 15x on cache hits. Two caveats, both from Tencent's own framing:
- Hy4 itself is currently on a 50% discount, so the real gap between the two models today is narrower than the list-price comparison suggests.
- The headline number comes from Tencent's internal WorkIndex benchmark, where Hy4 flash cost one-twentieth as much per task as the full Hy4. Tencent also says it cost one-eighth as much per task as DeepSeek's V4.1-Flash-1017.
Both figures are vendor-reported internal benchmarks, not independent evaluations. Treat them as a signal about relative cost, not a ranking of capability.
What the blind test says — and what it doesn't
Tencent also ran a blind test in which 85 WorkBuddy users rated model outputs across 259 tasks. According to Tencent, Hy4 flash scored close to Claude Haiku 5.5 in user satisfaction, while having the lowest cost and the shortest completion time among the models tested.
That is a useful data point, but note what it is: 85 users, 259 tasks, on Tencent's own product surface, rating satisfaction rather than a standardized capability benchmark. It tells you the model feels good enough for everyday office work. It does not tell you how it performs on hard reasoning, coding, or agentic benchmarks against the frontier.
This is the honest frame for Hy4 flash: it is a model optimized for "good enough, much cheaper, much faster" on routine work, not a new capability ceiling.
Where you can use it
- WorkBuddy — Tencent's office assistant, free access through November 30
- Yuanbao — Tencent's consumer AI app
- Tencent Cloud TokenHub — API access, currently at half the list price
- OpenRouter — planned, not yet live
- Weights — Tencent plans to release the model's weights, following the open-weights approach of the Hy4 flagship
The pattern worth watching
Hy4 flash is part of a clear trend among Chinese model labs: after the flagship race, the second race is the efficiency race. DeepSeek, Zhipu, and now Tencent are all shipping "flash" variants that trade a slice of capability for a large cut in cost and latency — and they are being judged on internal cost-per-task benchmarks rather than public leaderboards.
For builders, the practical question is not whether Hy4 flash beats the flagship. It doesn't, and Tencent isn't claiming it does. The question is whether the tasks you actually run — drafting, summarizing, editing spreadsheets, triaging long documents — are close enough in quality at one-twentieth the cost.
If your workload is mostly routine, that trade is worth testing. If your workload is hard reasoning or long-horizon agentic coding, the flagship — or a frontier closed model — is still the right comparison.
Hy4 flash is a bet that most daily AI work belongs in the first category. The next few weeks of real usage will show whether Tencent priced that bet right.