Hy4 Flash vs Hy4
Tencent's task recommendations alongside its listed prices and reported speed figures.
Pick Hy4 Flash for everyday work and Hy4 for long-horizon software engineering, according to Tencent. Tencent recommends Flash for document processing, deep search, slide generation, simple front ends and computer use. It recommends Hy4 when a complex engineering task needs Hy4's higher intelligence ceiling. On the Tencent Cloud TokenHub price page, the listed rates are US dollars (USD) 0.07 per million input tokens and USD 0.281 per million output tokens for Flash, versus USD 0.834 and USD 2.501 for Hy4.
Tencent reports that Flash decodes above 150 tokens per second and costs one-twentieth as much as Hy4 per task on Tencent's WorkIndex evaluation. Those are Tencent's measurements, not an independently established result. Tencent's announcement is in Chinese. The US-dollar rates appear on a separate TokenHub page. The two pages also use different currencies and describe limited-time discounts, so the listed rates should not be treated as converted prices or as confirmed post-discount charges.
Hy4 Flash vs. Hy4: Side-by-Side
| Dimension | Hy4 Flash | Hy4 |
|---|---|---|
| Total and active parameters | Tencent states 224 billion total parameters, including 64 billion Engram parameters, with 12 billion active parameters. | Tencent does not list it on its Hy site, Hy4 announcement or TokenHub page. |
| Context window | Tencent states a 1 million-token context length. | Tencent's Hy home page states a 1 million-token context window. |
| Input price in US dollars | USD 0.07 per million input tokens, as listed on the TokenHub page. | USD 0.834 per million input tokens, as listed on the TokenHub page. |
| Output price in US dollars | USD 0.281 per million output tokens, as listed on the TokenHub page. | USD 2.501 per million output tokens, as listed on the TokenHub page. |
| Cache price in US dollars | USD 0.003 per million tokens in the Cache column, as listed on the TokenHub page. | USD 0.042 per million tokens in the Cache column, as listed on the TokenHub page. |
| Prices in Chinese yuan | Tencent's Hy site lists Chinese yuan (CNY) 0.02 cached input, CNY 0.5 input and CNY 2 output per million tokens. | Tencent's Hy site lists CNY 0.3 cached input, CNY 6 input and CNY 18 output per million tokens. |
| Reported decoding speed | Tencent reports decoding above 150 tokens per second. | Tencent does not list it on its Hy site, Hy4 announcement or TokenHub page. |
| WorkIndex cost per task | Tencent reports one-twentieth of Hy4's cost per task on its internal WorkIndex evaluation. | Tencent uses Hy4 as the cost-per-task reference for the reported WorkIndex comparison. |
| Recommended tasks | Tencent recommends most everyday work, including deep search, document processing, slide generation, simple front ends and computer use. | Tencent recommends complex work such as long-horizon software engineering. |
| Access and release status | Tencent says Flash is live on TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima and Marvis. OpenRouter access and an open-source release are planned. | Tencent's Hy4 announcement is dated November 4, 2026 (Beijing time). |
| Limited-time discount end | Tencent states that its 50% inference-fee discount ends at 12:00:00 on November 30, 2026, Beijing Time. | Tencent states that its 50% inference-fee discount ends at 23:59:59 on December 31, 2026, Beijing Time. |
Are you one of these vendors? Update your listing
Hy4 Flash vs Hy4 for Everyday and Engineering Work
Hy4 Flash is Tencent's pick for most everyday tasks. Hy4 is its pick for complex long-horizon software engineering. Tencent gives task recommendations but does not publish Hy4's parameter count on its Hy site, Hy4 announcement or TokenHub page. Tencent publishes Flash's 224 billion total and 12 billion active parameters, but does not give a Hy4 parameter count on its Hy site, Hy4 announcement or TokenHub page. A parameter comparison would therefore leave one side blank.
For document processing or slide generation, start with Flash under Tencent's guidance. Deep search, a simple front-end display and computer use also appear in Tencent's Flash recommendations. If the task becomes sustained software engineering, keep Hy4 available rather than assuming that a lower token price will compensate for a weaker result. Tencent describes Hy4 as having a higher intelligence ceiling for that work.
The initial price choice is clear at the listed US-dollar rates: Flash has lower input, output and cache figures on the TokenHub page. The result for a particular job is less certain. TokenHub lists separate input, output and cache rates, while Tencent's per-task comparison comes from its own evaluation. Compare complete task outputs and usage on your samples before deciding which tasks should follow the cheaper route.
Hy4 Flash vs Hy4 Side-by-Side Prices
Tencent publishes separate US-dollar and Chinese-yuan price listings for Hy4 Flash and Hy4. The comparison table keeps those currencies apart. Its US-dollar figures are as listed on the TokenHub page. Tencent's Hy site supplies the Chinese-yuan figures. Neither listing establishes an exchange rate for comparing one page's numbers with the other's. The AI model price list provides a broader cross-model reference, but the Tencent pages supply the rates used here.
The TokenHub page labels one column Cache (USD / million tokens). It lists USD 0.003 for Flash and USD 0.042 for Hy4 in that column, alongside separate input and output rates. If your workflow reuses prompt material, record cached tokens separately when you compare requests. A single input-token rate cannot describe a request that also generates output and uses cached context.
Tencent's Hy site orders its Chinese-yuan prices as cached input, input and output. In that order, it lists CNY 0.02, CNY 0.5 and CNY 2 per million tokens for Flash. Hy4's corresponding listed rates are CNY 0.3, CNY 6 and CNY 18. Keep the units and order attached to each figure when you copy the prices into a budget: CNY 2 is Flash's output figure, while CNY 6 is Hy4's input figure.
TokenHub displays region tabs for Singapore, Guangzhou and Silicon Valley. The figures above are as listed on the TokenHub page. They have not been established separately for every tab. TokenHub also does not say whether its displayed US-dollar figures are before or after its limited-time discounts. Use the displayed rates as listed figures, without calculating a discounted US-dollar rate.
Tencent-Reported Cost and Speed
Tencent reports that Hy4 Flash costs one-twentieth as much as Hy4 per WorkIndex task. WorkIndex is Tencent's internal evaluation built from typical WorkBuddy tasks and human evaluation. Tencent says it examines output quality, single-turn output length and end-to-end time for everyday work. Its eight internal sets cover slide creation and revision, human-AI work on slides and Markdown, document editing, spreadsheet repair, audio and video processing, and visual understanding. The WorkIndex result concerns Tencent's listed everyday-work tasks, not every possible task.
Tencent also reports a one-thirteenth cost-per-task figure against Hy4 preview and a one-eighth figure against DeepSeek-V4.1-Flash-1017 at the stated max tier and busy-hour price. DeepSeek-V4.1-Flash-1017 is the October 17, 2026 update of the model covered in the linked DeepSeek V4.1 Flash guide, and Tencent specifies its max tier and busy-hour price. Keep the model version and Tencent's pricing conditions attached to that comparison.
For speed, Tencent reports Flash decoding above 150 tokens per second. A WorkBuddy product manager quoted in Tencent's announcement says average task time fell about 70% versus Hy4 and average cost was about one-fifteenth as high. The same manager says everyday results were basically equal, with a small drop on long-horizon coding tasks. A WeCom algorithm lead quoted by Tencent reports 88% faster output than Hy4 and about 60% fewer thinking-mode output tokens. These are product-team evaluations reported on Tencent's own page, rather than measurements from a shared external test.
Discount Dates and Task-Cost Limits
Tencent gives Hy4 Flash and Hy4 different discount end dates. The TokenHub pricing notice states: "Hy4 flash Limited-Time Discount Notes: From the date of this notice until 12:00:00 on November 30, 2026 (Beijing Time), Hy4 flash model inference fees are settled at 50% of the list price. The discount applies only to usage actually incurred during the above period. After the period expires, the model fees will be settled at the list price again."
The same notice states: "Hy4 Limited-Time Discount Notes: From the date of this notice until 23:59:59 on December 31, 2026 (Beijing Time), Hy4 model inference fees are settled at 50% of the list price. The discount applies only to usage actually incurred during the above period. After the period expires, the model fees will be settled at the list price again." The notices give different discount end dates for the two models.
Do not turn those notices into a quoted post-discount US-dollar price. TokenHub does not state whether the displayed dollar figures already include the discount, and it provides no separate post-discount dollar row in the cited source. Tencent's WorkIndex ratio is a per-task result from its evaluation, not a discount percentage. Likewise, the WorkBuddy manager's one-fifteenth cost estimate describes that product team's tasks. Neither ratio tells you the bill for a document job with your own input and output lengths.
Architecture and Task Fit
Tencent describes Hy4 Flash as a multimodal model with a 1 million-token context length. Its Hy home page also gives Hy4 a 1 million-token context window and describes Hy4's visual understanding and computer-use abilities. Equal published context lengths do not settle which model gives a better answer on a long job. Tencent's task guidance separates everyday work from complex engineering even though both listings state the same context length.
Tencent says Flash combines hybrid attention, LatentMoE (Tencent's term) and Engram. According to Tencent, Flash's key-value (KV) cache, which holds earlier tokens, is one-sixth the size of Hy4's. Tencent also describes self-distillation from Hy4 followed by reinforcement learning (RL) that cut end-to-end task output length by 50%. The reported reduction does not specify how many tokens your task will use.
Tencent attributes Flash's inference setup to a megakernel, 4-bit quantization and DFly optimizations. Tencent lists token rates and reports speed and task results. Hy4's page instead emphasizes long-horizon software engineering, cybersecurity research, professional financial analysis and complex game development. If one of those complex jobs is the main requirement, evaluate Hy4 on that job rather than treating Flash's faster reported decoding as a substitute for output quality.
When to Choose Each Model
Choose Hy4 Flash first when the work matches Tencent's everyday-task list and cost matters across repeated requests. A document-processing queue is a clear candidate because Tencent explicitly recommends Flash for document processing. Keep a record of the request, result and token usage for each sample. That gives you a way to compare a completed job instead of comparing input prices while overlooking generated output.
Choose Hy4 first for long-horizon software engineering. Tencent specifically directs that work to Hy4 and says the model has a higher intelligence ceiling. Its Hy home page also lists cybersecurity research, professional financial analysis and complex game development among Hy4's suitable tasks. The page does not publish a rule that every task in those fields requires Hy4. Send representative work to both models if you need to decide where an individual task belongs.
You can keep both models in a routing plan. TokenHub lists the application programming interface (API) parameter hy4-flash for Flash and hy4 for Hy4, so the two model choices have distinct identifiers. Put routine document jobs on the Flash path and reserve a Hy4 path for engineering samples. If Flash's answer fails a requirement that Hy4 meets, route that class of work to Hy4. If both meet the requirement, compare the recorded task cost rather than assuming Tencent's WorkIndex ratio will repeat on every request.
Who Should Not Rely on Either Recommendation
Teams that require independently reproduced performance figures should not adopt either model on Tencent's benchmark claims alone. Tencent reports the WorkIndex ratios, the blind test and the product-team speed figures. The cited material does not turn those results into independent evidence. Instead, use Tencent's guidance to choose which samples to send to each model, then judge the outputs against the requirements of your own work.
Tencent says its blind test involved 85 internal WorkBuddy users, 27 task categories and 259 everyday tasks. It reports satisfaction close to Claude Haiku 5.5, the lowest cost per task among models in that test, and the shortest task time. Those findings concern Tencent's everyday-task test. They do not answer whether Flash meets a particular engineering requirement. Tencent recommends Hy4 for long-horizon software engineering.
Tencent says it plans to open-source Flash. The announcement does not say that release has happened. Tencent says OpenRouter access is coming soon, while listing Flash as live on TokenHub and several Tencent products. If OpenRouter access or released weights are a requirement, wait for the relevant release rather than treating a plan as current availability.
What Would Change the Answer
Independent task results that contradict Tencent's everyday-work findings would weaken the case for routing routine work to Hy4 Flash. A smaller price gap would also change the cost decision, especially if both models produced acceptable outputs on the same jobs. Tencent's limited-time discounts end on different dates, so a decision based on charges during either discount period needs another look after that period ends.
A change in Tencent's task recommendations would affect the engineering boundary. For now, Tencent recommends Hy4 for complex long-horizon software engineering and Flash for most everyday work. Use those recommendations as a starting assignment, not as proof that every document or code task will finish at the reported WorkIndex cost. For a Hy4 Flash vs Hy4 decision, route routine document tasks to Flash, keep a Hy4 route for engineering work, and compare outputs on your own samples.
The Verdict
Pick Hy4 Flash for high-volume everyday work such as document processing and slide generation. Tencent recommends it for those jobs, lists lower input and output rates on TokenHub, and reports a WorkIndex cost per task one-twentieth of Hy4's. The task-cost result comes from Tencent's evaluation. Use your own completed jobs to decide whether the saving holds.
Pick Hy4 for complex long-horizon software engineering. Tencent says Hy4 has the higher intelligence ceiling for that work, and a WorkBuddy product manager quoted by Tencent reports a small Flash drop on long coding tasks. Do not assign an engineering job to Flash solely because Tencent reports faster decoding and lower everyday-task cost.
The Flash recommendation changes if independent results disagree with Tencent's reports, your samples favor Hy4, or the price gap closes. The Hy4 recommendation changes for a class of engineering tasks if Flash meets the same output requirements at a lower recorded cost. Tencent's reported figures establish a starting route. Your sample outputs determine whether to keep it.
Researched from primary vendor documentation and public regulator sources. Pricing and availability are accurate as of Nov 17, 2026 and can change. Confirm current terms with each vendor before you buy.
Frequently Asked Questions
- For everyday work, Tencent recommends Hy4 Flash and reports lower WorkIndex cost per task. For complex long-horizon software engineering, Tencent recommends Hy4 because it says Hy4 has a higher intelligence ceiling. Tencent's figures do not establish one winner for every task.
- Tencent reports that Flash costs one-twentieth as much per WorkIndex task, but that is an internal evaluation result. As listed on the TokenHub page, Flash costs USD 0.07 per million input tokens and USD 0.281 per million output tokens. Hy4 costs USD 0.834 and USD 2.501 respectively. TokenHub does not establish whether those displayed dollar rates include its limited-time discounts.
- Yes. TokenHub lists separate API parameters,
hy4-flashandhy4. You can assign Tencent's recommended everyday tasks to Flash and retain Hy4 for engineering work, then compare completed results before fixing a routing rule. - Tencent says it plans to open-source Hy4 Flash. Its November 16, 2026 announcement describes a planned release, not an available one. Tencent lists Flash as live on TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima and Marvis. It says OpenRouter access is coming soon.
Plan a Hy4 Model Routing Decision
At Layer3Labs, we build AI workflow automation for small and mid-sized teams. We can help define task samples and routing rules for a Hy4 Flash and Hy4 deployment.
Book a Consultation