A generated title card reading 'Hy4 flash' with the subtitle '224B total, 12B active, one million tokens of context', a price strip reading '$0.070 input / $0.281 output / $0.003 cached per million tokens', and a footer reading 'Released 16 November 2026. Prices are Tencent's list rates.'
Guides & Insights

Hy4 flash Lands at $0.070 an Input Million: Tencent's Cheapest Hy Model Is Also Its Least Documented

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

T​encent released Hy4 flash on 16 November 2026, and the specification reads like a cost-engineering exercise more than a capability one: 224 billion total parameters of which 64 billion are Engram parameters, 12 billion active per token, native multimodality, a 1,048,576-token context, and a mainland list price of ¥0.5 per million input tokens against ¥2 per million output. The company's framing is that it matches Claude Haiku 5.5 on everyday work in a blind test run on its own employees, at roughly an eighteenth of the per-task cost. The blind test is real and the price is real; the comparison across them is T​encent's own arithmetic, and the difference between those two statements is most of what anyone deploying this model needs to know.

There is no model card. There is no architecture paper, no licence, no weights repository and no independent benchmark run. What T​encent published instead is a research page with benchmark results baked into PNG charts, a set of internal product testimonials, and a promise to open-source the model later. That is a thinner dossier than Hy4 preview shipped with in August, and the gap matters more here than it would for a flagship, because the entire pitch for Hy4 flash is efficiency — and efficiency claims are the ones you cannot check from a price table.

What the rate card actually says

Per million tokens, on the mainland card: ¥0.5 input, ¥2 output, ¥0.02 cached input. T​encent quotes the same model in dollars as $0.070 input, $0.281 output and $0.003 cached, which is the figure the launch post led with. Set it beside the rest of the family and the positioning is unambiguous.

• Hy4 flash — ¥0.5 input / ¥2 output / ¥0.02 cached, 224B total with 12B active.

• T​encent HY4 — ¥6 input / ¥18 output / ¥0.3 cached, 770B total with 49B active, the general-availability build from 4 November 2026.

• Hy4 preview — the same ¥6 / ¥18 / ¥0.3 rate, Apache 2.0, weights downloadable since 28 August 2026.

That is a twelve-fold cut on input and nine-fold on output against its own flagship, from one generation to the next, with no change to the context window. Cached input falls by fifteen times, which is the line item that decides the economics of any agent that re-sends a long system prompt on every step. For a workload that reads a large document once and then asks twenty questions of it, the cache rate is the rate that matters, and $0.003 per million is priced to make that pattern nearly free.

Two promotions sit on top of the list price and both are temporary. T​encent Cloud's TokenHub is running a 50% discount for a limited period with no stated end date, and WorkBuddy is free to use until 30 November. Neither should be modelled into a production budget; the list rate is the number that survives December.

Where 12 billion active parameters come from

A generated diagram titled 'Four levers, one bill' with four labelled rows. Pre-training reads: hybrid attention, LatentMoE and Engram, giving a KV cache one twenty-fourth the size of Hy3's and one sixth of Hy4's. Post-training reads: self-distillation from HY4 followed by reinforcement learning, output length down 50 per cent. Inference reads: a megakernel, 4-bit weights and DFly, over 150 tokens per second. Disclosure reads: no layer count, no expert count, no model card. A footer reads 'All figures are Tencent's own, published 16 November 2026 without independent measurement.'

The parameter split is the first thing to understand, because 64 billion Engram parameters inside a 224-billion-parameter model is not a normal arrangement. Engram, in T​encent's usage, is a memory-style component that holds learned associations rather than participating in every forward pass the way an expert does. The practical consequence is that a large share of the model's total size is not charged against per-token compute: 12 billion parameters activate per token, which is a quarter of what GLM-5.3 Flash activates and roughly a quarter of T​encent HY4's 49 billion.

T​encent names four levers, and each one attacks a different part of the bill.

Pre-training. Hybrid attention combined with LatentMoE and the Engram parameters. T​encent's stated result is a KV cache one twenty-fourth the size of Hy3's and one sixth of Hy4's. On long-context serving that is the difference between a request fitting on a node and a request needing a cluster, and it is the single most consequential number in the release for anyone renting GPUs by the hour.

Post-training. Self-distillation from T​encent HY4, then reinforcement learning. The stated outcome is a 50% reduction in end-to-end output length. This is where a flash model usually gives back part of its price advantage — shorter answers can mean shallower work — and it is also where the per-task cost claims below get most of their leverage. Output tokens are the expensive ones at $0.281 per million, so halving them halves the larger side of the invoice before the price cut is even applied.

Inference. A megakernel, four-bit quantization and a decoding path T​encent calls DFly, together sustaining better than 150 tokens per second of decode. Fast decode changes agent behaviour as much as it changes user experience: a tool-calling loop that spends less wall-clock per step can afford more steps inside the same latency budget.

Architecture detail. None published. Layer count, expert count, routing width, the size of the vision encoder and whether multimodality is native or bolted on are all absent. The 224B and 12B figures come from the research page; everything finer-grained does not exist in public.

The blind test, and what "matches Claude Haiku 5.5" covers

The claim that travelled is the blind test. T​encent ran it with 85 users of its WorkBuddy product across 27 task categories and 259 real daily-work tasks, comparing Hy4 flash with several unnamed alternatives and reporting that user satisfaction came out close to Claude Haiku 5.5, that Hy4 flash had the lowest per-task cost of any model in the panel, and that it finished tasks in the shortest time.

Read the wording carefully, because the launch post and the research page do not say the same thing. The post says the model "matches" Claude Haiku 5.5. The page says satisfaction was close to it. Close is not equal, and a vendor summarising its own panel will pick the framing that flatters. More importantly, the panel is 85 employees of the company that built the model, working on tasks the company selected, in an interface the company also builds. That is a reasonable way to decide whether to ship a model internally. It is not evidence that a third party's mileage will match, and it does not establish parity on any task type the 27 categories happened not to include.

What the test does establish is narrower and still useful: on office-shaped work — documents, slides, spreadsheets, the sort of thing WorkBuddy is for — a 12B-active model was not obviously worse than a frontier-adjacent competitor, as judged by people who use both daily. If that generalises, the cost case is overwhelming. Whether it generalises is the open question, and there is no public benchmark in the release that would answer it.

Tracing the 18x

The "18x task cost reduction" in the launch post is not a price ratio. Hy4 flash costs twelve times less per input token than T​encent HY4, so a simple per-token comparison gives 12x, not 18x. The extra comes from T​encent's WorkIndex, an internal eight-suite benchmark covering PowerPoint creation, PowerPoint revision, human-machine co-writing, Markdown co-writing, document editing, table repair, audio and video processing, and visual understanding. On that suite, at the maximum tier and at busy-hour prices, T​encent reports per-task cost at one-thirteenth of Hy4 preview, one-twentieth of T​encent HY4, and one-eighth of DeepSeek-V4.1-Flash-1017.

One-twentieth against its own flagship and one-eighteenth against the model the post names is the same order of magnitude, so the 18x is a defensible round number — but it is a per-task figure, which means it folds in output length, retry rate and step count alongside price. Two of those three inputs are things T​encent controls through post-training and reports only as a percentage. A model that answers in half the tokens at a twelfth of the price is not 18x cheaper because of the price; it is 18x cheaper because it is also being made to talk less. If your task needs the longer answer, you get less than 18x, and the article you are reading cannot tell you how much less.

The same caveat applies to every per-task figure in the release, including the internal product numbers T​encent volunteered: a 70% cut in task time and roughly one-fifteenth the cost in WorkBuddy against T​encent HY4, a 2.52% general-chat gain against GLM-5.3-Flash in 元宝, 29% fewer output tokens and 6% fewer steps in 微信支付 agent flows against DeepSeek-V4.1-Flash-1017, 88% faster output and 60% fewer thinking tokens in 企业微信, and 22% to 39% fewer requests with 40% to 61% fewer tokens in 王者荣耀 root-cause analysis. These are the most credible numbers in the package, because they come from production systems with real users and real bills. They are also unaudited and selectively released, and no product publishes the flows where the model lost.

What the release does not contain

No open weights. T​encent says it plans to open-source Hy4 flash and to iterate on it frequently. A plan is not a licence. Hy4 preview shipped under Apache 2.0 on day one with a Hugging Face repository, vLLM and SGLang recipes and published sampling defaults; Hy4 flash has none of that, and the wording around the open-source commitment is softer than the preview's was.

No independent evaluation. As of publication, Artificial Analysis has not posted a score for Hy4 flash, and no third party has reproduced a single number from the release. Every figure above is T​encent's.

No machine-readable benchmark table. The DeepSWE, Terminal-Bench 4.0, APEX-Agents, AutomationBench, HLE and OSWorld 2.1 results exist as chart images on the vendor page, not as text. That is a strange choice for a launch whose whole argument is measurable efficiency, and it means the numbers cannot be copied, diffed or checked without reading pixels.

No safety or system card. Also no statement about refusal behaviour, data handling or regional availability beyond the mainland products named.

Where it runs, and how to try it cheaply

Availability on launch day is T​encent's own surface: TokenHub on T​encent Cloud for API access, WorkBuddy for office tasks free until 30 November, and 元宝, 企业微信, QQ, ima and Marvis on the consumer and enterprise side. Nothing else is announced.

A screenshot of the OrcaRouter model catalogue browse view, headed 226 models from 17 providers behind one API key and one bill, with filter facets for input modalities, context length, input price, status, access tier, series, supported parameters, privacy and inference region. Visible model cards include google/gemini-4-carbon at $2.00 input and $10.00 output per million tokens with a 1,048,576-token context, anthropic/claude-fable-5-2 at $8.00 and $40.00, minimax/minimax-m3.1-flash at $0.15 and $0.60, and deepseek/deepseek-v4.1-flash-1017 with a 1M-token context. No Hy4 flash card appears in the listing.

On OrcaRouter, Hy4 flash is not in the catalogue yet — the route that exists for this family is tencent/hy4-preview, at $0.834 per million input and $2.501 per million output, alongside a rate-limited free tier. That is the model Hy4 flash was distilled from, so it is the right baseline to measure against while you wait: run the same tasks on both once the flash route lands and you will have the per-task ratio that T​encent's WorkIndex implies without having to trust it.

The reason to care about where it is served is the price model. OrcaRouter passes provider list price through at 0% markup, which is not a discount but a commitment about what happens when a vendor cuts one: T​encent has already run two promotional discounts on this model in its first week, and a pass-through route reflects the next one the day it appears rather than when a reseller updates a table. Failover matters for the same reason in reverse — an unproven 12B-active model on launch-week infrastructure is exactly the model you want to be able to fall back from without rewriting a client, and the same key reaches more than 200 models if it does.

What to point it at first

T​encent's own deployment guidance splits the family cleanly, which is more useful than most vendor advice because it concedes a weakness: Hy4 flash for deep search, document processing, PowerPoint generation, simple frontend work and computer use; T​encent HY4 for long-horizon software engineering. The concession is in the last clause — the cheap model is not being sold as a coding replacement for the expensive one, and the internal WorkBuddy data shows a small drop on long coding tasks even as it reaches parity on daily work.

A generated decision card titled 'Point it at the short tasks' with three rows: work the vendor names for Hy4 flash reads deep search, documents, slides, simple frontend and computer use; work it names for Tencent HY4 reads long-horizon software engineering; and a measurement row reads twenty real tasks with input and output tokens, cache hits and wall-clock recorded. A footer reads 'Deployment guidance is Tencent's; the cost ratios are per-task and unaudited.'

That maps onto a simple test plan. Take twenty tasks you already run, of which most are document or agent-flavoured and a few are genuinely hard engineering ones. Record input tokens, output tokens, cache hits, wall-clock and whether the output was usable. Run them on Hy4 flash and on the baseline you use today. You will learn in an afternoon whether the 18x applies to your traffic, which is more than the release itself will tell you.

The cheapest Hy model is the least legible one

T​encent has built something worth paying attention to: a quarter-active model with a million-token window, a KV cache small enough to change the serving maths, and a price that undercuts its own flagship by an order of magnitude. The engineering described — Engram parameters, a 50% output-length cut through self-distillation, four-bit inference at over 150 tokens per second — is the kind of work that shows up in a bill rather than a leaderboard, which is presumably why T​encent led with per-task cost.

But the release asks to be believed on exactly the quantities it declines to document. No weights, no model card, no independent score, benchmark numbers behind pixels, and a headline comparison drawn from a panel of its own employees. None of that is dishonest. All of it means the honest summary is "this is probably the best value in the Hy family and nobody outside T​encent has verified it," and the right response is a two-week trial with token counts, not a migration.

The open-source promise is the thing to diary. If T​encent ships weights the way it did for Hy4 preview — Apache 2.0, day one, with serving recipes — the efficiency claims become checkable and the model becomes a serious option for anyone who wants to run cheap inference themselves. Until that happens, Hy4 flash is a very good price attached to a very thin dossier, and the difference between those two facts is where your evaluation budget should go.

Compared in this article3

Detected from this article · Benchmarks: Artificial Analysis · updated daily