AI DevelopmentNew Release9 min readPublished

Tencent Hy4 Flash: Open Weights, 1M Context and Licence

The licence distinguishes a product that uses the model from a business selling access to it. The hardware and API bills make a second distinction between owning the weights and renting inference.

DA
Digital Applied
PublishedNovember 19, 2026
Read time9 min

A company with $60 million in sales wants to add a button that extracts tables from customers’ invoices. Another company with the same revenue wants to sell developers a general-purpose model API. Under the Tencent Hy License, those two products can face different commercial requirements even if both run identical Hy4 Flash weights. The first question is what customers can do with the model. Annual revenue comes next.

Tencent announced Hy4 Flash on November 16. Developers can now download the instruction-tuned checkpoint or call a hosted version with text and image inputs. It generates text and supports a 1,048,576-token context. The checkpoint card describes about 12 billion active parameters per token, but the standard weight files occupy about 440.5 GB. For most teams evaluating a document workflow, a few API calls are a cheaper way to find out whether the answers are usable. Tencent’s release announcement · Tencent’s checkpoint card

Key takeaways
  1. 01
    Identify a model-service business before applying the size testsSection 10 covers large operators that offer model inference or fine-tuning to others. It expressly excludes certain feature-limited products and mere relays.
  2. 02
    The licence grants worldwide rights subject to its conditionsThe released Flash licence names no excluded territory. Revenue is measured worldwide. The separate rules of a hosting provider still apply to customers buying its service.
  3. 03
    Active weights govern compute more directly than storageThe model card breaks the approximately 220B checkpoint into a language backbone, an Engram lookup table, a vision encoder and a speculative drafter. The published FP8 variant is still about 287.1 GB.
  4. 04
    Price the hosted route at both the introductory and standard rateTokenHub shows $0.035 per million input tokens and $0.1405 per million output tokens during its 50% promotion. Its billing notice places the cutoff at noon Beijing time on November 30.

01 — Model IDsWhat exactly are you running?

On Hugging Face, tencent/Hy4-flash is the instruct checkpoint. Tencent also publishes tencent/Hy4-flash-FP8, a lower-precision instruct version for serving, and tencent/Hy4-flash-base, a pretrained base checkpoint. If the task is answering a user’s question about a receipt, start with an instruct model. The base release is chiefly relevant to teams that want to develop or adapt a model and can evaluate the result themselves.

The hosted names differ. Tencent Cloud TokenHub’s request parameter is hy4-flash. OpenRouter lists the route as tencent/hy4-flash. A repository name identifies files you can download. A router name identifies a service that takes a request and returns output. If you compare these routes, record the checkpoint revision or provider, the date and the serving settings. The same family name gives little insight into quantization, time spent in a queue or how much conversation history an endpoint will accept. Tencent Cloud model details · OpenRouter model listing · The earlier Hy4 release

02 — Legal scopeWho actually crosses the threshold?

The file that accompanies the weights is the Tencent Hy License, Version 1.0. It derives from Apache 2.0 and adds its own commercial conditions. Section 2 grants a worldwide copyright licence subject to those conditions. There is no list of excluded countries or regions in the released Flash licence. The conditions travel with the checkpoint, including when a team modifies or quantizes the weights. Hosted service availability and account terms are separate matters.

Start with section 10.1. Its definition of model-as-a-service, or MaaS, covers giving other parties remote access to inference or fine-tuning, where they can supply inputs, parameters or training data. Two exclusions matter. One is an end-user product that confines model access to particular features or workflows and offers no general model access. The other is a service that merely passes requests to an independently operated third-party model service. A network call inside an application does not, by itself, classify the application.

For an operator of a MaaS business, section 10.2 requires a separate written licence from Tencent if either size test is met. The revenue test uses aggregate worldwide gross revenue above US$50 million or its currency equivalent in any consecutive twelve months. It counts revenue from all sources. The user test applies before using the model in or for a product or service the operator provides with more than 100 million monthly active users in any calendar month. These are thresholds the measurement must exceed. The licence’s definition of the legal entity includes entities under common control, which also matters when scoping the revenue calculation.

Section 10.3 exempts solely internal use from section 10.2 when neither the weights, their derivatives nor their outputs are supplied to third parties. Section 10.4 directs licensing questions to hy_partners@tencent.com. The general redistribution duties in section 4 still matter when you distribute a derivative. They include providing the licence, identifying modified files and keeping the applicable notices. Tencent Hy License attached to the checkpoint reviewed November 19

03 — Product designFollow the request to its user

Consider the invoice feature in the opening example. Customers upload an invoice and receive predefined fields. If they cannot send general prompts, change inference settings or use the endpoint as a model service, the feature fits the licence’s narrow-workflow exclusion on those facts. Its company revenue alone does not trigger section 10.2. Giving those same customers a developer key they can use to submit arbitrary model requests changes the product you need to classify.

Employees summarize internal reports
Model, derivatives and outputs stay within the organization. Section 10.3 specifically exempts this internal use from section 10.2. Revisit the premise before sending outputs to outside clients.
Internal use
A business sells a general inference endpoint
A $60 million worldwide gross-revenue operator serving its own Hy4 Flash deployment crosses the revenue test even if it earned only $10,000 from that endpoint. It needs Tencent’s written licence before using or continuing to use the model.
MaaS operator
A gateway only forwards calls to another service
Section 10.1 excludes merely relaying a request to an independently operated third-party model service. Running your own inference as a fallback changes the facts for that route.
Relay exclusion

A large enterprise can operate several of these arrangements at once. Section 10.2 starts by asking whether the operator runs a MaaS business. If it does, evaluate the obligations of that operator across its uses as well as the features of the proposed product. In a design review, draw the boundary for each route. Which legal entity serves the weights? What can the external user submit? Who gets a copy of the output? That account is more useful than a yes-or-no label on an internal model inventory. An application that starts as a single-purpose reporting tool can acquire general-purpose access in a later release. How model-specific licences change deployment decisions

04 — Model designWhere do the other weights go?

Hy4 Flash uses a mixture of experts. Of its 42 backbone layers, 41 contain routed experts. Each token uses 16 of 512 such experts alongside a shared expert. The release also includes Engram, which looks up learned representations of token pairs and triples. Tencent’s checkpoint card assigns 155.5B parameters to the language backbone, 63.5B to Engram, 0.43B to the vision encoder and 0.85B to the speculative drafter. It reports 11.3B active text parameters and another 0.43B for vision. Active work can be much smaller than the collection of weights needed across different requests.

Tencent’s corporate announcement says 224B total, while the checkpoint card describes 220B. Its four listed components add to about 220.28B using the rounded figures. Use the published files to plan a deployment. Either total describes a model much bigger than a dense 12B download. It is perfectly possible to spend less compute generating the next token and still need a multi-accelerator server and substantial host memory to make the weights available. Active parameters, cost and speed

Twelve attention layers select sparse context, while thirty use sliding windows. The five-layer DFly2 drafter proposes upcoming tokens and reuses the target model’s stored attention information. Tencent reports a key-value cache one-sixth the size of Hy4’s and one twenty-fourth the size of Hy3’s. That cache holds information from the current sequence. Its reduction helps explain why a long session can become cheaper to serve. It does not shrink the checkpoint files by the same factor.

05 — Serving capacityA million tokens on which machine?

The instruct checkpoint’s weight shards total about 440.5 GB on disk. The published FP8 checkpoint is about 287.1 GB. FP8 changes the representation of many parameters, while some tensors retain higher precision. A simple one-byte-per-parameter estimate misses that difference. The vLLM project uses the download sizes multiplied by 1.2 to provide aggregate memory planning hints of 529 GB and 345 GB respectively. It explicitly labels these estimates. In its implementation Engram tables can sit in host RAM, so a disk size or a rule-of-thumb multiplier cannot establish a minimum GPU memory requirement.

vLLM’s default recipe uses eight H200 GPUs for its BF16 configuration. It starts with an 8,192-token limit, four concurrent sequences and one image per prompt. The recipe identifies Hopper and Blackwell data-center GPUs and a dedicated container image. Tencent’s own examples run the FP8 checkpoint with tensor parallelism across eight devices in either vLLM or SGLang. None of these examples establishes the smallest viable configuration at one million tokens. vLLM’s Hy4 Flash recipe and memory assumptions · Tencent’s FP8 checkpoint

Suppose your product sends a 6,000-token report most of the time and a 400,000-token bundle once a day. Measure time to first answer and memory with several ordinary requests running together. Then introduce one large request and repeat. Processing the long input before decoding begins can affect another user’s wait even when the first user’s answer is brief. On TokenHub, the provider separately advertises 960K maximum input and 64K maximum output within a 1,024K context. OpenRouter lists 1,048,576 context and 65,536 maximum output. Inventory the limit for the route you will operate. Comparing weight sizes and hardware requirements

06 — API economicsThe hosted bill after November

On November 19 Tencent Cloud showed a 50% introductory discount. Its model card gives the US-dollar prices below. Tencent’s online-inference billing notice specifies that eligible usage receives the discount until noon Beijing time on November 30, 2026, with standard billing afterward. OpenRouter’s November 19 route listing matched the three discounted rates. Use the dated router price for a current comparison and confirm its own listing when forecasting a later month. Tencent Cloud billing notice · TokenHub’s USD model price card

US dollars per million tokens, read November 19, 2026. Cached input and uncached input are separate quantities. The example excludes tool fees and retries.
Hy4 Flash routeUncached inputCached inputOutput
TokenHub promotional$0.035$0.0015$0.1405
TokenHub standard$0.070$0.003$0.281
OpenRouter, November 19$0.035$0.0015$0.1405

Consider 10,000 document jobs. Each uses 8,000 uncached input tokens, 2,000 cached input tokens and 1,000 billed output tokens. On the promotional rate, 80 million uncached tokens cost $2.80, 20 million cache hits cost $0.03 and 10 million output tokens cost $1.405. Total model spend is $4.235, or about $4.24. Standard rates put the same quantities at $8.47. These figures assume the advertised cache hits actually occur and that the stated output includes all billed generation.

At that volume, a self-hosted case needs a reason beyond avoiding the token bill. Keeping images in a controlled environment, customizing the model or operating with stable high utilization can be such reasons. Add the cost of idle hours, redundancy and someone operating the server. Then divide by completed jobs at your required quality. A lower per-token bill loses its value quickly if a human has to fix missing invoice lines.

07 — Vendor evaluationWhere does the speed claim hold?

Tencent reports decoding above 150 tokens per second with its optimized kernels, four-bit quantization and DFly2 speculative decoding. Its downloadable FP8 checkpoint and the eight-device example use a different stated precision. Decode speed covers the output phase. It gives no stand-alone prediction for how long a request spends uploading an image, processing a long prompt or waiting for a tool response. Tencent also attributes a 50% reduction in average end-to-end output tokens to reinforcement learning. Treat that as a result from its training and evaluation setup.

Tencent’s internal WorkIndex measures eight sets of office and agent tasks. Flash scored 42.1 against 39.7 for Hy4 Preview and 51.4 for Hy4. Tencent says the average cost per task was one thirteenth of Preview’s, one twentieth of Hy4’s and one eighth of DeepSeek-V4.1-Flash-1017 at its Max setting and peak-hour pricing. The cost comparison and the quality score point to different decisions. A team that can accept Flash’s result on routine edits may gain from its price. A task that needs Hy4’s additional capability can absorb repeated attempts on the cheap model.

The benchmark appendix shows why the average deserves unpacking. On Tencent’s run of APEX-Agents v1.1, Flash scored 63.1 to Hy4 Preview’s 53.9. On Terminal-Bench 4.0 the order reversed, with 19.2 for Flash and 26.3 for Preview. The appendix describes different agent programs and turn budgets across tests, and evaluates each model at its highest available reasoning setting. For an invoice extractor, a trial with your own awkward spreadsheets and scanned documents is more informative than either score in isolation. Tencent’s benchmark appendix and evaluation notes

A practical trial

Keep one document with a split table, one with an ambiguous total and one with a chart whose labels live inside an image. Decide which fields must match the source before sending any request. Compare accepted outputs, reviewer minutes and billed tokens. Use the longest realistic input as a separate capacity test.

Deployment decision

Choose the service boundary before buying hardware

For a defined document feature, start with the current licence and a representative hosted trial. For an external model API, establish the section 10 position before rolling out access. For a self-hosted service, build the budget from the checkpoint, the length of a real request and the number of people waiting at the same time. Our AI transformation services cover workflow evaluation and deployment planning.

Digital Applied

Evaluate an open-weight model on a real workflow

Compare hosted cost, hardware needs and task acceptance before choosing a deployment.

Model evaluationHosting economicsDocument workflows
Evaluation inputs

Start with an actual request

  • →A representative document
  • →The fields a reviewer would accept
  • →Peak request volume and context length
Answers

Frequently asked questions

It uses the Tencent Hy License, Version 1.0. The file retains provisions derived from Apache 2.0 and adds commercial conditions specific to large model-service businesses. Use the licence accompanying the revision you deploy.

Digital Applied newsletter

Deep dives on AI, marketing and development.

Practical guides and fresh insights by email. No recycled takes.

Related dispatches

Related reading

AI Development

Mistral Large 4 Goes Open: Price vs DeepSeek, Kimi, GLM

Mistral released Large 4 weights on October 29 and lists $0.70 input and $2.10 output per million tokens. Direct prices compared with DeepSeek, Kimi and GLM.

October 29, 2026 · 12 minRead
AI Development

Microsoft MAI Realtime 1: $0.01 Audio Minutes Plus Reasoning

Price Microsoft’s MAI Realtime 1 with its separate reasoning fees, compare voice API billing units and see what teams can test in the limited Foundry preview.

November 19, 2026 · 8 minRead
AI Development

Gemini 4 Neon vs GPT-6.2 Sol vs Claude Sonnet 5.5 Pricing

Compare direct API prices, the 272K-token cutoff, cache costs and independent coding scores. See the bill for 100 jobs on Gemini 4 Neon, GPT-6.2 Sol and Sonnet.

November 19, 2026 · 9 minRead
AI Development

Microsoft MAI-Image-2.7 and Flash: Editing, Cost, Access

Compare Microsoft’s two image editors on price per 1,000 outputs, editing votes, reference images and commercial terms for MAI Labs, Azure Foundry and fal.

November 19, 2026 · 8 minRead
AI Development

AI Search Agents Compared: Google, Perplexity, ChatGPT

Google's always-on information agents, Perplexity Pro, and ChatGPT Search compared. Which AI search agent delivers the best research results in 2026?

May 20, 2026 · 14 minRead
AI Development

OpenAI + Dell Codex: On-Premises Enterprise Agents

OpenAI and Dell partner to bring Codex to hybrid and on-premises environments via Dell AI Factory. What changes for enterprise coding workflows.

May 18, 2026 · 12 minRead
Google Search

See more Digital Applied analysis in your Google results by adding us as a preferred source.

Add as a preferred source