THE SIGNAL IN ONE SENTENCE
A cheap token is a fine thing. It is also an excellent way to lose sight of the bill. Tencent released Hy4 flash on November 16 as the high-volume member of its Hy4 model family. The company is positioning it for daily work such as document processing, presentations, spreadsheet repair, research, simple front-end development, image understanding and computer use. The headline numbers are built for a procurement slide. Hy4 flash has 224 billion total parameters, including 64 billion Engram parameters, while activating 12 billion parameters for a token. It accepts text and images, supports a context window of 1,024K tokens, allows up to 960K input tokens and up to 64K output tokens through Tencent Cloud TokenHub. The standard API price listed by Tencent is about $0.070 per million input tokens, $0.281 per million output tokens and $0.003 per million cached input tokens. TokenHub is advertising a 50 percent launch discount through November 30. That is genuinely inexpensive access to a large multimodal model. It still does not tell you what the work costs. A model can be cheap per token and expensive per accepted task. It can produce an answer quickly, then spend three retries repairing a spreadsheet formula. It can draft a presentation for pennies, then require an analyst to rebuild the charts. It can read a long document cheaply while quietly missing the clause that determines the entire decision. The useful unit is not one million tokens. It is one completed task that passes a clear acceptance test. Start with what changed. Hy4 flash uses a mixture-of-experts design. Tencent says its pretraining architecture combines hybrid attention, LatentMoE and Engram features. The company reports that its key-value cache is one twenty-fourth the size of Hy3's and one sixth the size of Hy4's. That cache matters because a model serving a long conversation has to retain information needed to continue generating. Smaller cache requirements can let an operator hold more active work or spend less memory per session. Tencent also says reinforcement learning reduced average end-to-end task output length by 50 percent. On the serving side, it credits a megakernel, four-bit quantization and a technique called DFly for reported decoding speeds above 150 tokens per second. These are Tencent's engineering claims. They help explain the product direction. They are not a promise about the speed, memory or quality of your workload. The API is concrete enough to test now. TokenHub lists the model ID as hy4-flash. It supports OpenAI Chat Completions, OpenAI Responses and Anthropic Messages protocols. The service lists deep reasoning, structured output, function calling, prompt caching and image understanding among the supported capabilities. That compatibility lowers migration friction. It does not eliminate behavioral differences. Changing one model string can preserve the shape of a request while changing the answer length, tool selection, JSON reliability, image interpretation, refusal behavior, language quality and tendency to keep working after the task is already done. So build a task ledger before moving production traffic. Choose a representative set of jobs, not a bag of trivia questions. For a document workflow, include summarization, clause extraction, revision, citation checking and a case where the model should admit that the source is incomplete. For spreadsheets, include broken formulas, hidden sheets, date formats, merged cells, charts, cross-sheet references and a workbook that must remain unchanged outside the target range. For presentations, include a written brief, a source file, brand constraints, a chart with known values and an accessibility check. Judge the actual deck, not the prose describing the deck. For a computer-use agent, include login boundaries, unexpected dialogs, ambiguous buttons, network failures, permission requests and a stop condition that requires a person. Then define accepted. An accepted document may need every factual claim linked to the supplied source, no invented numbers, correct formatting and no unresolved placeholders. An accepted spreadsheet may need formulas preserved, totals reconciled, the target error corrected and a machine-readable change log. An accepted presentation may need accurate data, usable hierarchy, editable elements, the right aspect ratio and no text outside the safe area. If the acceptance test is fuzzy, the cost comparison will be fuzzy too. Now count the entire task. Record uncached input tokens, cached input tokens, output tokens, reasoning tokens if billed or exposed, image inputs, tool calls, failed tool calls, retries, wall-clock time and human review time. Include the cost of any search, browser, storage, sandbox or third-party service used by the agent. For each task, calculate four numbers. First, raw model cost: the tokens and media charged by the provider. Second, run cost: raw model cost plus tools and infrastructure. Third, attempted-task cost: run cost across the first attempt and every retry. Fourth, accepted-task cost: all attempted-task cost plus the human time required to inspect and repair the result, divided by the number of outputs that passed. That last number is the one worth putting on the procurement slide. Suppose Model A costs one cent per run and passes 50 percent of the time. Model B costs four cents and passes 90 percent. If a failed attempt needs a five-minute human inspection, Model A may be the expensive option before anyone opens the invoice. Cheap generation can create costly review. This is especially important for Hy4 flash because Tencent's own launch argument is based on task economics. The company created an internal benchmark called WorkIndex around typical WorkBuddy activities. Tencent says the evaluation covers eight sets involving slide creation, slide editing and human-AI co-writing, Markdown co-writing, document editing, spreadsheet repair, audio and video processing, and visual understanding. It also reports seven public agent evaluations: DeepSWE, ProgramBench, Terminal-Bench 4.0, APEX-Agents, AutomationBench, HLE and OSWorld 2.1. Tencent says Hy4 flash slightly outperformed the much larger Hy4 Preview in overall effectiveness. On its internal WorkIndex, it reports per-task cost equal to one thirteenth of Hy4 Preview, one twentieth of Hy4 and one eighth of DeepSeek-V4.1-Flash-1017 under the specified Max-tier peak-hour pricing. Those ratios are interesting. They are also vendor-run comparisons whose exact task distribution, scoring details, retry policy and operational prices may differ from yours. Tencent separately reports a blind evaluation with 85 internal WorkBuddy users across 259 tasks in 27 categories. It says Hy4 flash achieved user satisfaction close to Claude Haiku 5.5 while posting lower task cost and the shortest task-completion time among the compared models. That is a sensible direction for evaluation because people judged work instead of only benchmark answers. But internal users know the product. Their prompts, patience and expectations may not resemble a customer, attorney, designer, finance analyst or support agent encountering the system in the wild. Recreate the comparison with your own people and files. Blind the reviewers when practical. Remove the model name, price and expected winner. Randomize output order. Give every model the same source materials, tools, time limit and acceptance criteria. Keep the failures, not just the attractive outputs. Measure quality by task family. A model that is excellent at summarizing Chinese documents may be weaker at repairing a financial workbook in English. Strong visual understanding does not guarantee exact extraction from a dense table. Fast decoding does not guarantee fast task completion if the agent takes a wrong turn and keeps going at 150 tokens per second. Speed can make a mistake arrive sooner. Long context deserves the same skepticism. A one-million-token window means the service accepts a very large combined history. It does not mean every detail receives equal attention, that facts at the middle are reliably recovered or that filling the window is economical. Test long context with planted facts at different positions, conflicting versions, irrelevant attachments and repeated instructions. Ask the model to cite exact source locations. Measure whether quality changes as the prompt grows. Use the prompt cache deliberately. If many tasks reuse the same policy manual, style guide or schema, caching can lower repeated input cost. Record the cache hit rate instead of assuming the lowest listed price applies. A cache price is not useful when every request changes the supposedly reusable prefix. Version the prefix too. When a policy file changes, the team should know which tasks used the older cached material and which used the replacement. For agent workflows, inspect tool behavior separately from language quality. Track whether Hy4 flash selects the right function, supplies valid arguments, preserves required identifiers, waits for a result and stops when the tool fails. Check structured outputs against the schema before another system consumes them. Do not let a low token price turn retries into background noise. Set a maximum attempt count, maximum token budget, maximum tool budget and maximum wall-clock time for each task. Require a clear failure result when the budget ends. An agent that loops cheaply is still looping. Human review should also have a budget and a purpose. For low-risk drafting, a quick inspection may be enough. For financial, legal, personnel, medical, security or public-facing work, define which facts a qualified person must verify and which actions remain prohibited without approval. Measure reviewer disagreement. If two reviewers regularly disagree about whether the output passes, the acceptance rule needs repair before the model comparison can mean much. Geography and language matter here. Tencent built Hy4 flash inside a Chinese product ecosystem and positions it across WorkBuddy, Yuanbao, WeCom, QQ, ima and Marvis. That can make it especially relevant for Chinese-language work and organizations already using Tencent services. Global teams should test every intended language, locale, date format, currency, writing style and regulatory context. A presentation that sounds natural in one language may become stiff or misleading after localization. A spreadsheet repair that assumes one decimal convention can damage another. Data governance belongs in the price test too. Before sending production files, document the service region, data terms, retention, access controls, encryption, subprocessors and incident path. Count the engineering work needed to redact sensitive fields, isolate credentials, log tool activity and meet internal policy. The cheapest permitted model is more useful than the cheapest model on a public price table. Finally, keep the release status honest. Hy4 flash is hosted at launch. Tencent says it plans to open-source the model and iterate frequently, but that future release is not available today. Do not write deployment plans around weights that have not been published. Evaluate the hosted service that exists. Hy4 flash may prove to be an unusually economical daily model. The architecture is aimed at using less active compute, the API prices are low, the context window is large and the service supports the protocols and tools many teams already use. That is enough reason to run a serious trial. It is not enough reason to multiply tokens by price and call the answer total cost. Pick real tasks. Define accepted. Count retries, tools, latency and human repair. Separate easy work from consequential work. Test languages, files and long context. Cap the agent's budget. Recheck after every model update. The winning model is not the one that speaks for the fewest pennies. It is the one that gets useful work across the finish line with the least total waste.
01
WHAT ACTUALLY CHANGED
Tencent released Hy4 flash on November 16, 2026 as a lower-cost member of the Hy4 model family for high-volume daily work.
Tencent says the model has 224 billion total parameters, including 64 billion Engram parameters, with 12 billion active parameters.
The model accepts text and images and supports a 1,024K-token context window through Tencent Cloud TokenHub.
TokenHub documents up to 960K input tokens and 64K output tokens for the hy4-flash model ID.
TokenHub lists support for deep reasoning, structured output, function calling, prompt caching and image understanding.
The service supports OpenAI Chat Completions, OpenAI Responses and Anthropic Messages protocols.
Tencent lists standard prices of about $0.070 per million input tokens, $0.281 per million output tokens and $0.003 per million cached input tokens.
Tencent says Hy4 flash exceeds 150 tokens per second under its serving setup and reduces average end-to-end task output length by 50 percent.
The model is available through TokenHub and several Tencent products, with OpenRouter integration announced as coming soon.
Tencent says it plans to open-source Hy4 flash, but the launch is hosted access rather than an open-weight release.
02
WHY THIS MATTERS
Low API prices can make multimodal and agent workflows practical at volumes where larger models are too expensive.
Token price alone excludes retries, tool charges, infrastructure, latency, review and repair.
A smaller active parameter count can reduce inference work without requiring a small total model.
Protocol compatibility can lower migration effort while leaving model behavior, tool use and output quality different.
A one-million-token context window expands possible inputs but does not prove reliable attention across the full window.
Prompt caching can reduce repeated input cost only when the reusable prefix actually produces cache hits.
Vendor task-cost ratios are more relevant than isolated benchmark scores but still need independent reproduction on local workloads.
Fast decoding can increase wasted output when an agent follows the wrong plan or loops through tools.
Human repair time can reverse the apparent price advantage between models.
Hosted availability and future open-source plans are different deployment states with different governance choices.
03
WHERE IT COULD HELP
- Build a representative acceptance set from real documents, spreadsheets, presentations, images and computer-use tasks.
- Define a pass condition for every task before comparing models.
- Track uncached input, cached input, output, media, tools, retries and wall-clock time per attempt.
- Add reviewer time and repair time to the cost of each accepted deliverable.
- Compare cost per accepted task instead of cost per isolated run.
- Blind and randomize human evaluations when practical so model names and prices do not influence ratings.
- Separate results by language, task family, risk level and file type.
- Plant facts at different positions in long prompts and require exact source citations.
- Test conflicting versions and irrelevant attachments before trusting the large context window.
- Measure prompt-cache hit rate and version every reusable prefix.
- Validate structured output against its schema before another system consumes it.
- Test function selection, argument validity, identifier preservation and stopping behavior separately.
- Set maximum tokens, attempts, tool calls, wall-clock time and external spend for every agent task.
- Require a clear failure result when the task budget ends.
- Keep human approval for consequential financial, legal, personnel, medical, security and public actions.
- Record reviewer disagreement and refine ambiguous acceptance rules.
- Test languages, currencies, date formats and local conventions used by the actual audience.
- Verify service region, retention, access control, encryption and incident handling before sending sensitive files.
- Evaluate the currently hosted service instead of assuming promised open weights are available.
- Repeat the acceptance set after every model, price, tool or serving update.
KEEP A HAND ON THE WHEEL
The architecture, speed, cache-size, output-length, benchmark, task-cost and blind-evaluation results are reported by Tencent. They describe Tencent's selected serving stack, internal WorkIndex tasks, public benchmark setup and WorkBuddy participants. They are not independent guarantees for another workload or deployment. Tencent reports a blind evaluation with 85 internal users across 259 tasks in 27 categories. Internal users may differ from external customers in language, familiarity, task selection and quality expectations. The one-million-token context figure is a capacity limit, not evidence that every fact remains equally retrievable across the full window. The reported decoding speed does not include every tool, network, review or repair step in an agent workflow. Hy4 flash is available as a hosted model at launch. Tencent says open-source availability is planned, so local weight access, license terms, quantization artifacts and self-hosted performance remain future questions. Watch for independent task-level comparisons, exact WorkIndex methodology, error and retry distributions, multilingual results, model-version behavior, stable OpenRouter availability, published weights, license terms and evidence that low token prices translate into lower total cost per accepted result.
04
TERMS WORTH KEEPING
OPEN GLOSSARY CARD
Mixture of experts
A model architecture with many specialist components that activates only a selected few for each input.
OPEN GLOSSARY CARD
Context window
The amount of input and generated material a model can consider during one call.
OPEN GLOSSARY CARD
Context window
The amount of information a model can actively consider at one time.
SOURCES AND VERIFICATION STATUS
This article was written from the materials below. Product claims and dates were checked against those sources on November 17, 2026.
THE PUBLICATION ENGINE
WANT A SIGNAL OF YOUR OWN?
We build source-grounded publications, private briefings, and editorial systems for organizations with something useful to say.
WORK WITH US