Tencent Launches Hy4 Flash at 92% Cheaper Input Tokens Than Hy4
Tencent's Hy4 flash ships with 224B parameters, a 1M-token window and prices of $0.070/$0.281 per million tokens — roughly 92% and 89% below full Hy4.

Why it matters
- Hy4 flash launched November 16, 2026 at $0.070 per million input tokens, about 92% below full Hy4's $0.834.
- The model has 224 billion total parameters but activates 12 billion per token, with a one-million-token context window.
- Tencent reports a key-value cache one-sixth the size of Hy4's and decoding above 150 tokens per second.
- Blind testing by 85 WorkBuddy users across 259 tasks showed satisfaction close to Claude Haiku 5.5 at the lowest cost per task.
- WorkBuddy access is free and TokenHub API usage is 50% off through November 30, 2026; an open-weight release is planned.
Tencent launched Hy4 flash on November 16, 2026, a natively multimodal model that cuts input token prices by roughly 92% and output token prices by about 89% compared with the full Hy4 model released twelve days earlier. The move turns Tencent's Hy4 line into a deliberate two-tier offering: a cheap, fast model for high-volume office work, and the flagship for harder jobs.
The pricing gap is stark. Tencent Cloud's price list puts Hy4 flash at $0.070 per million input tokens and $0.281 per million output tokens. Full Hy4 charges $0.834 and $2.501 for the same volumes. Cached input on Hy4 flash costs $0.003 per million tokens, against $0.042 on Hy4 — a 93% reduction on the cheapest tier of usage.
Why does that matter? Because inference cost, not raw capability, increasingly decides which model companies route everyday work through. If Hy4 flash holds up outside Tencent's own benchmarks, it lands in the same price band where OpenAI, Google and DeepSeek fight for high-volume productivity and agent traffic — the largest single source of API demand.
How Tencent squeezed the costs
Hy4 flash is a big model with a small activation footprint. It contains 224 billion total parameters, including 64 billion Engram parameters, but activates only 12 billion parameters per token. That sparse-activation design is the first lever: you pay for the full model once when serving it, but each request runs through a fraction of it.
The second lever is memory. Tencent reports that its hybrid-attention, LatentMoE and Engram design shrinks the key-value cache — the store of earlier context computations the model reuses as it reads — to one twenty-fourth the size of Hy3's cache and one-sixth of Hy4's. A smaller cache means more concurrent users per GPU, which translates directly into lower per-token prices.
Post-training adds a third lever. Hy4 flash was refined through self-distillation from full Hy4, a teacher-student setup where the smaller model learns to mimic the larger one. Reinforcement learning during this phase cut average end-to-end output tokens by 50%, meaning the model answers in half the generated text.
On the serving side, Tencent credits megakernels, four-bit quantization and a system called DFly for pushing decoding speed beyond 150 tokens per second. The takeaway: the price cut comes from architecture, caching and output compression working together, not from one trick.
What is WorkIndex, and why did Tencent build it?
Tencent chose not to lead with public benchmark scores. Instead it built WorkIndex, an evaluation suite modeled on tasks from WorkBuddy, its office assistant product. WorkIndex measures three things that public leaderboards mostly ignore: deliverable quality, single-turn output length and task turnaround time.
The suite spans eight evaluation sets:
- Slide creation and editing
- Human-AI slide co-writing
- Human-AI Markdown co-writing
- Document editing
- Spreadsheet repair
- Audio and video processing
- Visual understanding
That framing signals where Tencent thinks the enterprise market is heading. Buyers care whether the model produces a finished slide deck or a repaired spreadsheet quickly and cheaply, not whether it adds a few points on an abstract reasoning test.
What do the internal results show?
On WorkIndex, Tencent reports that Hy4 flash scores slightly above Hy4 preview overall, while its average cost per task falls to one-thirteenth of Hy4 preview's, one-twentieth of full Hy4's, and one-eighth of DeepSeek-V4.1-Flash-1017's at Max-tier peak-hour pricing. A model that is marginally better on office deliverables at a twentieth of the cost is a different product category, not a small iteration.
Tencent also ran a blind evaluation with 85 internal WorkBuddy users who completed 259 tasks across 27 categories. Participants reported satisfaction close to Claude Haiku 5.5, along with the shortest end-to-end completion time and the lowest cost per task among the evaluated models. The caveats deserve emphasis: these are Tencent's own users, Tencent's own tasks and Tencent's own reporting.
Where can developers get it, and what does the API support?
Hy4 flash is available now through Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima and Marvis, with OpenRouter availability planned. Two promotions run through November 30, 2026: WorkBuddy access is free until that date, and TokenHub API usage carries a 50% discount until 12:00 Beijing time that day. The discount applies only to usage generated during the promotional window.
The API identifier is hy4-flash. Tencent Cloud's documentation lists the following capabilities:
- Image understanding
- Reasoning
- Structured output
- Function calling
- Caching
Maximum input is 960,000 tokens and maximum output is 64,000 tokens — both below the model's one-million-token context window, leaving headroom for system prompts and context. The API supports OpenAI Chat Completions, OpenAI Responses and Anthropic Messages formats, which lowers switching costs for developers already integrated with those ecosystems. Customers must activate the model service before making calls.
Hy4 flash or full Hy4: which one for which job?
Tencent draws the line clearly. It recommends Hy4 flash for deep search, document processing, slide creation, simple frontend demonstrations and computer use. It continues to recommend full Hy4 for complex, long-horizon software engineering, because the flagship retains the higher intelligence ceiling.
In other words, flash is for volume; the flagship is for depth. That two-tier structure mirrors how rivals have split their lineups, but Tencent's split is unusually aggressive on price.
The release completes a plan Tencent announced on November 4, 2026, when it introduced full Hy4 with visual understanding and computer-use capabilities and said a more cost-efficient Flash version would follow soon. Hy4 flash is that follow-through, arriving within two weeks.
Tencent also says it plans an open-weight Hy4 flash release and frequent updates. The open-weight promise is the detail worth watching: if independent developers can reproduce the 150-tokens-per-second serving speeds and one-twentieth task costs on their own hardware, Hy4 flash stops being a Tencent Cloud discount and becomes genuine pressure on every incumbent pricing sheet.
Source: unite.ai
More from Sophie Lindqvist
Show full bio
Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.
749 articles
Related articles
- Tencent ships Hy4 with vision and computer use, claims second place in China
- DeepSeek V4.1-Flash-1017 Outscores Larger Rivals on Agentic Benchmarks
- GPT-6.1 Terra halves API prices and closes the gap on OpenAI's biggest models
- OpenAI ships GPT-6.2 Astra and Sol with 20% cheaper API pricing
- Google Launches Gemini 3.1 Flash-Lite Starting at $0.25 Per Million Tokens