NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

Tencent Promised 20x. The Price Sheet Stops at 12.

Tencent Holdings (0700.HK) lists two rows on its Hunyuan site. Hy4 flash: 0.02, 0.5, 2.

2026-11-17 · 4842 words · NeuroAI
Tencent Promised 20x. The Price Sheet Stops at 12.

Tencent Holdings (0700.HK) lists two rows on its Hunyuan site. Hy4 flash: 0.02, 0.5, 2. Hy4: 0.3, 6, 18. Per million tokens, in the order cache hit, input, output.

The launch post concludes: flash costs 1/20 of Hy4 per task.

Run the table. Input 6÷0.5 is 12 times. Output 18÷2 is 9 times. Cache 0.3÷0.02 is 15 times. At equal token counts the best case is 1/12, the worst 1/9, and the more output-heavy the task, the closer it sits to 9.

The 20 cannot come out of the price sheet.

It sits in token count. For 1/20 to hold, flash must finish a task on 0.45 to 0.6 times the tokens Hy4 burns: 20÷12 is 1.67, inverted 0.6; 20÷9 is 2.22, inverted 0.45.

Put another way, 40% to 55% fewer tokens.

Tencent says as much: reinforcement learning cut "end-to-end task output length by 50%," which lands mid-range. Unit price 9 to 12 times cheaper, times roughly 2 times fewer tokens, arrives at 20. Tencent published the result, not the decomposition.

The same price, two different multiples

Hy4 production and Hy4 preview cost the same. Tencent Cloud's November 4 page says pricing carries over from preview — 6 yuan input, 18 yuan output, 0.3 yuan cache — more capability at the same price.

Yet flash is 1/13 against preview and 1/20 against production. Same denominator, two multiples. 20÷13 is 1.54.

Read it backwards: production burns 1.54 times the tokens preview does per task. The stronger model costs more because it runs longer, not because its tokens are dearer. You are paying for the extra steps it thinks through.

A second check against DeepSeek

The third multiple is 1/8, against DeepSeek-V4.1-Flash-1017 at the Max tier and peak-hour price.

That sheet lists 2 yuan input and 8 yuan output; flash charges 0.5 and 2. Both exactly one quarter. 1/8 divided by 1/4 is 1/2.

Two rivals, two unit-price gaps — 9 to 12 times against its own flagship, 4 times against DeepSeek — and the token layer lands on 2 times either way. The 50% sentence is the third number in the same ledger.

The 20 is behavioural, not a property of the product. If flash's output ever returned to Hy4's length, the claim would collapse to 1/12 or 1/9.

How a 12B model beats a 49B one

Flash carries 224B total parameters, 64B of them Engram, with 12B activated per token. Hy4 is 770B total and 49B active; preview matches at 770B and 49B. Tencent says the 12B model is "slightly better" than the 49B preview on aggregate quality.

Under a third of the parameters, under a quarter of the active ones, still ahead. Three things explain it. Pretraining adds hybrid attention, LatentMoE and Engram, with a KV cache 1/24 of Hy3's and 1/6 of Hy4's. Post-training self-distills from Hy4, then halves output length with reinforcement learning. Inference runs megakernel, 4bit quantization and DFly, decoding above 150 tps.

Only the last is pure engineering. The first two reshape the model — one shrinks the cache, the other shortens what it says.

The 64B Engram block comes with a number and no explanation of how it loads at inference, or whether it counts against VRAM.

Hy4 production shipped November 4, less than two weeks ago. Flash is chasing that model, not its predecessor.

Eight test sets, 259 tasks

Tencent skipped the public leaderboards and built WorkIndex, 8 internal eval sets: slide creation and editing, human-machine co-writing in slides and in Markdown, document editing, spreadsheet repair, audio and video processing, visual understanding.

Human evaluation used 85 internal WorkBuddy users, 27 categories and 259 tasks, blind. Three findings: satisfaction close to Claude Haiku 5.5, per-task cost below every model compared, shortest duration.

The suite measures artifact quality, single-turn output length and end-to-end time — not the ceiling of intelligence. "Slightly better than preview" applies to those 8 kinds of work. Tencent itself recommends Hy4 for long-horizon software engineering, an admission that flash wins on cost per unit of work, not capability.

259 tasks across 27 categories is under 10 each; only the aggregate holds.

Tencent also ran 7 public benchmarks — DeepSWE, ProgramBench, Terminal-Bench 4.0, APEX-Agents, AutomationBench, HLE, OSWorld 2.1 — but rested its case on the internal sets, arguing public boards sit "far from daily work." The price is that nobody outside can reproduce the 259 tasks.

What the discount is made of

KV cache at 1/6 saves memory and bandwidth. Halved output saves decode time. 4bit quantization plus megakernel saves compute per token. The 0.5 yuan is a result, not a cause.

Per-task cost is an average over one task set, one input-output mix, one retry profile. Change the set and the multiple moves.

Document teams land near 12. Code completion teams, with more back-and-forth and longer output, drift toward 9. Neither gets 20.

The price before November 30

Flash is live on Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima and Marvis, with OpenRouter coming. Until November 30, WorkBuddy is free for a limited time and TokenHub's API is 50% off: 0.25 yuan input, 1 yuan output, 0.01 yuan cache.

The Hunyuan team plans to open-source flash and iterate quickly. Released weights let third-party providers undercut 0.5 yuan — a road Hy4 has walked before, with revenue from cloud and products.

The promotional rate ends November 30; until then only the token question is live.

What is worth remembering is not 20 but its structure: a 9-to-12-times price gap and a 2-times token gap. The first moves with price wars; the second depends on how the model works.

For anyone building a budget: treat 20 as a ceiling, not a forecast. Measure tokens per task, multiply by unit price. The saving will mostly sit in prompts and retry counts.

*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*

Related coverage

Cite this article

NeuroAI (脑机网), "Tencent Promised 20x. The Price Sheet Stops at 12," 17 November 2026, https://neuroai.site/a/na-model-hy4-flash-price-1117

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Foundation Models” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories