Reported 16 Nov 2026 · Labs A one-minute read

Tencent launched a model at 0.5 yuan per million input tokens, weights not published.

Hy4 Flash went live on Tencent Cloud and in QQ and WeCom on November 16, Beijing afternoon. The caveat: the cost claims rest on Tencent's own benchmark, scored by Tencent.

Until now

Cheap Chinese models are smaller, weaker models sold on price.

Now

Tencent says a cheaper model beat its own larger one on its tests.

Why it matters

12 billion active parameters against 49 billion in the preview it is said to beat, with a cache one sixth the size and decoding above 150 tokens a second. The price is the product. Nothing is downloadable, so the cheapest way to run it is Tencent's cloud.

Established

  • Tencent published a price list: 0.5 yuan input, 2 yuan output per million tokens.
  • It is live on Tencent Cloud, WorkBuddy, Yuanbao, WeCom, QQ, ima and Marvis.
  • 224 billion total parameters, 12 billion active, 1 million token context.

Not established

  • That it beats rivals. WorkIndex is Tencent's own suite, scored by Tencent.
  • That weights follow. One tracker says open sourcing is planned, Tencent's coverage does not.
  • Independent quality. No outside leaderboard has ranked it.
The five-minute readThe record, the timeline, the weight and 5 sources

The record

Reported, not verified. Tencent's pages at hunyuan.tencent.com and cloud.tencent.com render in JavaScript and returned no readable text, so we could not read the model card or price list at source. The figures come from Science and Technology Daily, a Chinese state newspaper, and a model tracker citing Tencent's changelog. Verification event: reading Tencent's own Hy4 Flash page or the TokenHub price list directly, by December 17.

Tencent's Hunyuan team released Hy4 Flash on November 16, 2026, in the Beijing afternoon. The earliest timestamp we can fix is 17173.com at 14:31:33 Beijing, 06:31 Universal Time, inside this window; Science and Technology Daily, China's science ministry paper, filed at 21:08 Beijing. The morning window shut at 13:07 Beijing. It is a mixture-of-experts design: 224 billion total parameters, 64 billion of them what Tencent calls Engram parameters, 12 billion active per token, a context of 1,048,576 tokens, multimodal input. Pricing on Tencent Cloud TokenHub, per million tokens: 0.5 yuan input, 2 yuan output, 0.02 yuan cache hits, halved to 0.25, 1 and 0.01 under a promotion ending November 30 at 12:00 Beijing. Free on WorkBuddy until November 30. Tencent's efficiency claims: the key-value cache is one twenty-fourth of Hy3's and one sixth of Hy4's; decoding exceeds 150 tokens per second through megakernel work and 4-bit quantisation; reinforcement learning cut task output length by 50%. On WorkIndex, Tencent's internal suite of 8 test sets, the composite is 42.1 against 39.7 for Hy4 preview and 51.4 for the Hy4 flagship, and Tencent puts per-task cost at one thirteenth of Hy4 preview, one twentieth of Hy4 and one eighth of DeepSeek-V4.1-Flash at peak pricing. Hy4 preview has 770 billion total and 49 billion active parameters, so Hy4 Flash is credited with beating it on a quarter of the active parameters. In a blind test of 85 WorkBuddy users across 259 tasks in 27 categories, satisfaction was 2.79 of 4, against 2.81 for Claude Haiku 5.5, 3.09 for DeepSeek-V4.1-Flash and 2.53 for GLM-5.3-Flash. Public benchmark scores reported by one tracker include Terminal-Bench 4.0 at 19.2 and ProgramBench at 16.0, against the flagship's 44.4 and 30.0, plus DeepSWE 64.6, OSWorld 2.1 68.1 and HLE 55.2. No weights, no licence, no independent ranking.

How it got here

  1. Oct 6

    Mistral launches a 1 trillion parameter model and promises the weights that month. Edition 044. The promise is the pattern this card belongs to.

  2. Oct 14

    Tencent prices $1.8 billion and 26 billion yuan of bonds, paying 6.10% for 10-year dollars. Edition 050, the only Tencent entry in the ledger before today.

  3. Nov 15

    Huawei Cloud lists openPangu-2.1-Pro as open source with no weights, no parameter count and a price list, and retires its genuinely open 2.0 models. Edition 095.

  4. Nov 16

    Alibaba puts Qwen4-Max on sale in the Beijing morning, 71 billion active parameters against 95 billion, and says the same work costs 58% less. Edition 096, this morning.

  5. Nov 16

    Tencent launches Hy4 Flash in the Beijing afternoon, 12 billion active parameters, 0.5 yuan per million input tokens, nothing to download. The third Chinese flagship in 3 days without weights.

Who has to react

DeepSeek, named in Tencent's own cost comparison, and Zhipu, whose GLM-5.3-Flash lost the blind test; Alibaba, 11 hours ahead of it; Anthropic, whose Haiku 5.5 is the benchmark; and every Chinese developer who treated open weights as the local alternative to a metered API.

What to watch

  • Would strengthen itA fourth Chinese laboratory among DeepSeek, Zhipu, Moonshot, ByteDance or MiniMax ships a flagship as a paid endpoint with no downloadable weights, by March 31, 2027.
  • Would weaken itTencent publishes downloadable Hy4 Flash weights with a licence on Hugging Face, ModelScope or its own site, by March 31, 2027.
  • Would strengthen itAn independent leaderboard such as LMArena, Artificial Analysis or SWE-bench ranks Hy4 Flash at or above Hy4 preview, by March 31, 2027.

Why it weighs this much

Signal strength
CostMedium
CommitmentMedium
ReversibilityEasy
PersistenceYears

Cost is medium. Training a 224 billion parameter model and serving it across 7 products is expensive, and the promotion gives away half the revenue for 2 weeks, but the architecture makes serving cheap. Commitment is medium: a published price list, a dated promotion with an hour on it, and live availability across Tencent's own products are awkward to reverse quietly, though a model listing is not a contract. Reversibility is easy, which is the point of selling access rather than publishing weights: Tencent can reprice, restrict or withdraw this at will, as Huawei did to its 2.0 models on November 15. Persistence is years, not structural: the pricing resets and the next model replaces it.

The words

Active parameters
The share of a model's weights actually used for each token. Fewer active parameters means less computing per answer, so a lower price.
Mixture of experts
A design that splits the model into parts and wakes only a few per token, which is how 224 billion parameters cost like 12 billion.
Key-value cache
Memory holding what the model has already read. A smaller cache serves more users per chip, which shows up in price.
Open weights
Model files anyone can download and run themselves. Without them, the model exists only where its owner serves it.

Sources

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.

8 contributions. Thank you.

You have built
a record here.

Keep it: create an account to verify your contributions and put your name to the answers you choose. Your record stays yours.

Keep my record ↗
Optional: help show whose perspectives are missing
Kept with your answers under a random id, never with your name. Used for one thing: showing you, on your record, where you stand next to people like you, once at least 20 of them have answered.

Add your signature

Stand behind your judgment.

Create an account to attach your name to this answer. Your other answers show no name unless you sign them too.

Sign up to add your name ↗

Already have an account?