Tencent Hunyuan released Hy4flash on November 16, and the pitch is unusually narrow for a flagship-adjacent launch: it is not trying to be the smartest model, it is trying to be the cheapest one you can run all day.

The shape of the model

Hy4flash carries 224 billion total parameters, of which 64 billion are Engram parameters, and activates 12 billion per token. It handles multimodal input and a context window of one million tokens. Three design choices do the cost work: hybrid attention, LatentMoE routing, and the Engram memory layer.

The most consequential is KV cache compression. Tencent says the cache is one sixth the size of Hy4's, and decoding runs above 150 tokens per second. Smaller cache means more concurrent sessions per GPU, which is the difference between a model you demo and a model you leave running.

The price ladder

Per million tokens: ¥0.5 in, ¥2 out, ¥0.02 on a cache hit. That last number is the one to stare at. A cache hit costs four percent of a fresh input token, which is the kind of ratio that changes architecture decisions rather than just invoices.

Tencent's own comparison: Hy4flash is slightly better than the 770-billion-parameter Hy4preview in aggregate quality, while costing about one thirteenth per completed task. Against standard Hy4 the ratio is roughly one twentieth.

Where Tencent says to use it

The company is explicit about the split. Deep search, document processing, PPT generation and similar daily work should go to Hy4flash; long-horizon, complex tasks should stay on Hy4. That is a high-low pairing, and it is the same routing pattern every serious API buyer is converging on: cheap model for the bulk, expensive model for the hard tail.

The risk in that pattern is routing itself. If you route by task type you will get it wrong at the boundary, and the boundary is where most of the cost sits. Routing by retry — run flash, escalate on a failed check — is usually more reliable, at the cost of one extra call on the hard cases.

What is missing from the release

No independent benchmark sheet, and no published numbers on how quality degrades as context approaches the million-token limit. The 150 tokens per second figure is decode speed under Tencent's conditions; batch size and concurrency will move it.

Tencent says the model will be open-sourced later. Until then, the only way to price it is on your own traffic.

A practical test

Take one document-heavy workflow and one search-heavy workflow. Run each on Hy4flash with caching enabled and log the cache hit ratio. If hits stay above 50%, the economics are roughly what Tencent claims. If they sit near zero, you are paying ¥0.5 in and ¥2 out like any other mid-tier model, and the 12-billion active parameter ceiling will show up as quality loss on the tasks that need reasoning rather than extraction.