GPT-Image 2.6 Flare & Sunburst are liveImages from $0.025 each

Hy4 Flash: 1M context, price and API access

Tencent's Hy4 Flash is now both a hosted API model and a public weight release with roughly 220B total parameters under the Tencent Hy License.

Kiki, OmniaKey COO8 min readUpdated Nov 19, 2026
On this page

Hy4 Flash works two ways: as a hosted model on Tencent Cloud, OpenRouter and Vercel AI Gateway, and as public weights on Hugging Face. The card lists about 220 billion total parameters and 12 billion active per token under the Tencent Hy License 1.0 (checked November 19, 2026).

  • Hosted: hy4-flash on Tencent Cloud TokenHub; tencent/hy4-flash on OpenRouter and Vercel AI Gateway.
  • Weights: instruction-tuned, FP8 and Base repositories, all public and ungated, at roughly 220B total and 12B active.
  • Price: CNY 0.5 input and CNY 2 output per million tokens at list; a limited-time 50% rate halves both.

Updated November 19, 2026: the weights are now public. Prices, limits and repository details below were checked that day.

Can you use Hy4 Flash today?

Yes. Tencent Cloud lists hy4-flash as online, both gateways list tencent/hy4-flash, and the three Hugging Face repositories are public and ungated.

WhereModel IDStatus
Tencent Cloud TokenHubhy4-flashOnline
OpenRoutertencent/hy4-flashListed
Vercel AI Gatewaytencent/hy4-flashListed
Hugging Facetencent/Hy4-flashPublic

Tencent Cloud and both gateway listings give November 16, 2026 as the hosted release date. OpenRouter's canonical slug is tencent/hy4-flash-20261116. TokenHub takes three request styles under one model ID: OpenAI chat completions, the Responses API and Anthropic Messages, so OpenAI or Anthropic SDKs can reach the same model.

Which Hy4 Flash weights should you download?

Start with tencent/Hy4-flash unless you are serving at volume or fine-tuning. It is the instruction-tuned model with image input and tool calling. The FP8 repository is the same instruct model quantized for serving, and the Base repository is pretrained rather than chat-tuned.

RepositoryUse it forSize
Hy4-flashChat, vision, tools441 GB
Hy4-flash-FP8FP8 serving287 GB
Hy4-flash-baseFine-tuning439 GB

Sizes are repository storage totals rounded to the nearest gigabyte, not a memory requirement.

All three share an architecture: 42 layers, 512 routed experts with 16 active per token plus one shared expert, a 120,832-token vocabulary and a 1,048,576-token position limit. Text and images go in; text comes out. The card splits that into 155.5B language-model parameters outside Engram, vision and the draft model, plus 63.5B Engram, 0.43B vision and 0.85B DFly2. Those four parts add to 220.28B, and the weight files total 220.25B. Active work per token is 11.3B text parameters plus 0.43B vision, rounded to 12B.

Tencent's original November 16 launch material gave 224B; the latest weight card gives approximately 220B. Both counts are official; for capacity planning, use the card's approximate 220B.

The FP8 repo uses E4M3 weights in 128-by-128 blocks with dynamic activations, leaving the language-model head, token embeddings and hc_head unconverted.

What does the Tencent Hy License require?

The weights are downloadable and redistributable with attribution, under a custom license derived from Apache 2.0. Section 10 adds commercial terms for large model-as-a-service operators, so these are open weights under Tencent's custom terms; the file states that it is not Apache 2.0.

Redistribution requires a copy of the license, prominent notices on modified files and retention of attribution notices; your own terms cannot remove or limit Section 10. Fine-tuned, quantized, distilled, merged or pruned weights count as derivative works, and Section 10 travels with them.

Section 10 covers model-as-a-service: giving third parties remote access to inference or fine-tuning, including through an API. It requires a separate written license if the aggregate worldwide gross revenue of the entity and its affiliates, from any source, exceeds US$50 million in any 12 consecutive months, or if the relevant service exceeds 100 million monthly active users in a calendar month. Narrow end-user features without general-purpose model access, and services that merely relay requests to an independently operated third-party model, fall outside that definition. Internal-only use is exempt only if the work, its derivatives and their outputs stay away from third parties. Tencent's licensing contact is hy_partners@tencent.com.

The full text is in the Tencent Hy License 1.0. Hugging Face tags the repositories license: other with the name tencent-hy-license, and the file states that it is not Apache 2.0.

What hardware and software do Tencent's examples use?

Tencent publishes no GPU model, VRAM figure or minimum GPU count. What the card does give is two serving recipes, both using tensor parallel 8, plus a Transformers loading path.

The vLLM image is vllm/vllm-openai:hy4-flash. It loads the FP8 repository with --tensor-parallel-size 8, the FLASHMLA_SPARSE attention backend and a five-token DFly speculative config, and enables the hy_v4 tool-call and reasoning parsers. The SGLang image is lmsysorg/sglang:hy4-flash, run with --tp 8, DFLY at block size 6, on x86 or Arm. Both are published commands for an eight-GPU topology.

Transformers support is not in a release yet. The card says it will merge into the main branch soon and points to a pull-request commit, with AutoModelForImageTextToText and AutoProcessor as the loading classes.

Treat tensor parallel 8 as the example's layout, not a stated minimum or a promise about memory. The 287 GB FP8 download is a disk and transfer number. Size the machine against the checkpoint and your sequence length, then test your own traffic.

What are the context and API limits?

The context window is 1,048,576 tokens on the model card and in the config. Hosts write it differently, and their output caps do not match either.

PlaceContextOutput cap
TokenHub1024K64k
OpenRouter1,048,57665,536
Vercel1,024,00064,000
Card1,048,576none given

TokenHub caps input at 960k tokens. The card gives no output limit, and the three hosts differ by as much as 1,536 tokens. For long jobs, leave headroom under the host's cap rather than assuming the largest number applies everywhere.

Tencent Cloud's model page lists 300 requests and 10,000,000 tokens per minute; the page labels the request figure as maximum RPM. That is the first-party listing for TokenHub. The page also shows a 1,000,000-token free allowance with no expiry.

Input is text and images. Video and audio do not appear as supported inputs on the card or any of the three host listings. Reasoning accepts high (the default, thinking on) or no_think, and the card recommends temperature 0.9 with top_p 1.0.

What do hosted Hy4 Flash calls cost?

Tencent Cloud's model page lists CNY 0.5 per million input tokens, CNY 2 per million output tokens and CNY 0.02 per million cached input tokens. A limited-time 50% rate takes those to CNY 0.25, CNY 1 and CNY 0.01; the page shows a 0.5 discount on all three lines and does not state an end date.

Per 1M tokensInputOutputCache hit
List¥0.5¥2¥0.02
50% rate¥0.25¥1¥0.01

A request with 20,000 input tokens and 2,000 output tokens costs 20,000 × 0.5 / 1,000,000 + 2,000 × 2 / 1,000,000 = CNY 0.014 at list, or CNY 0.007 at the discounted rate. A hundred calls like that in a day come to CNY 1.4, or CNY 0.7 at the current rate. Reusing an identical prompt prefix bills the repeated part at CNY 0.02 per million instead of CNY 0.5, so put fixed instructions and shared documents first.

OpenRouter and Vercel AI Gateway both list dollar rates: 0.035 per million input tokens, 0.1405 per million output tokens and 0.0015 per million cached tokens, checked November 19. The same 20,000/2,000 request costs 20,000 × 0.035 / 1,000,000 + 2,000 × 0.1405 / 1,000,000 = $0.000981, or about a tenth of a cent. A hundred calls come to $0.0981. Tencent bills TokenHub in CNY, so keep the two lists separate.

Should you download the weights or call the API?

Call the hosted model when demand is spiky, when you want the 1M window without running your own serving stack, or when you need tool calls and structured output today. Download the weights when prompts and outputs must stay inside your own environment, when you want the Base checkpoint for fine-tuning, or when steady traffic is large enough to justify the hardware.

The license is the other deciding factor. If you plan to expose the model to third parties through an API, check Section 10 against the revenue and user thresholds for your entity and its affiliates before you build the service.

For another path to Tencent's models, OmniaKey's catalogue lists Hy4, the flagship, with current rates on the pricing page. Hy4 Flash is a separate model.

How does Hy4 Flash compare with Hy4?

Hy4 Flash sits below Hy4 in Tencent's line. Tencent's own WorkIndex score is 42.1 for Flash against 51.4 for Hy4, and the company puts Flash at 1/20 of Hy4's cost per task on the same internal set. It also claims decode speeds above 150 tokens per second with megakernels, 4-bit quantization and DFly2.

Those are vendor numbers, and we did not rerun them; use them to shortlist, then test your own tasks. The Hy4 release write-up covers what changed on November 4, and the coding and tool-call comparison of Hy4 against DeepSeek V4.1 Flash shows how the flagship behaved on small tasks. For current access and rates, see the Hy4 model page.

Frequently asked questions

Are the Hy4 Flash weights open?

Yes. The weights are public and redistributable under the Tencent Hy License 1.0, a custom license derived from Apache 2.0 with extra commercial terms in Section 10. It is not Apache 2.0, and Hugging Face tags it license: other.

How much does Hy4 Flash cost?

Tencent Cloud lists CNY 0.5 per million input tokens, CNY 2 per million output tokens and CNY 0.02 per million cached input tokens. A limited-time 50% rate halves those figures; OpenRouter and Vercel list 0.035, 0.1405 and 0.0015 per million tokens.

What are the Hy4 Flash context and output limits?

The card and config give a 1,048,576-token context. TokenHub caps output at 64k and input at 960k, OpenRouter at 65,536 output tokens and Vercel at 64,000; the card states no output cap.

Can I run Hy4 Flash on my own machine?

Tencent publishes no GPU minimum. Its examples use tensor parallel 8 for both vLLM and SGLang, and the FP8 download is 287 GB, so plan around a multi-GPU server and test the setup yourself.

Can I use the weights without a Tencent Cloud account?

Yes. The three Hugging Face repositories are public and ungated, and the license permits local use and redistribution with attribution. Hosted calls need a TokenHub, OpenRouter or Vercel account.

Sources checked

Updated November 19, 2026. We will revise the limits and prices when Tencent publishes an output cap or changes the TokenHub rate.

Hy4 FlashTencent Hunyuanlong-context AImultimodal modelsAPI access

Related articles

Recent posts

Popular guides