Tencent Opens Hy4 flash: 12B Active Parameters, 1M Context, and a Licence Clause Big API Hosts Need to Read First
TL;DR: the short version
Tencent's Hunyuan team announced Hy4 flash on November 17 and posted the weights to Hugging Face, ModelScope and GitCode on November 18 — three repositories in all: Hy4 flash (instruction-tuned), Hy4 flash-FP8 (quantised), and Hy4 flash-base (pretrained base). The model card and repo timestamps show the repos were created on November 17 and the weight files finished uploading on November 18.
Three things define what the model is trying to do:
- A very large model that wakes up only a small part of itself: roughly 220B total parameters, about 12B activated per token.
- A 1M-token context window, with native text and image input.
- Aggressive pricing: $0.07 per million input tokens, $0.003 per million cached input tokens.
What you should know before downloading: this is not Apache 2.0. It ships under the Tencent Hy License, an Apache 2.0 derivative with an added Section 10. If you run a model-as-a-service business with more than $50 million in annual revenue, or a product with more than 100 million monthly active users, you need a separate written licence from Tencent first.
What it is: a big, sparse, multimodal model
According to the official model card, Hy4 flash is assembled from four parts:
- A 42-layer language backbone: one dense feed-forward layer plus 41 mixture-of-experts layers, hidden size 3,072, vocabulary 120,832.
- A vision encoder: adapted from SigLIP2-So400m-Patch16-Res512, connected to the backbone through an MLP projector. Input is text or images; output is text.
- Engram (n-gram memory): an extra 63.5B parameters that store learned bigram and trigram representations, retrieved by multi-head hash lookups and injected at layers 1, 11, 21 and 31.
- DFly2, a drafter: five layers, 0.85B parameters, used for speculative decoding.
The parameter count is reported two ways, and they do not quite match. The press release says 224B total including 64B of Engram parameters; the model card table lists 155.5B of backbone, 63.5B of Engram, 0.43B of vision and 0.85B of drafter, which adds up to roughly 220B. The few-billion gap does not change the picture, but it is worth remembering when you compare numbers across papers and leaderboards.
Activated per token: 11.3B for text plus 0.43B for vision.
The context window is 1,048,576 tokens.
Why it is cheap: the savings sit in the KV cache and the output length
Tencent describes three engineering moves, and two of them are things a user can actually feel:
First, the KV cache shrank. The backbone interleaves sparse attention (IndexPool) with sliding-window attention at a 1:3 ratio, and routed experts compute in a lower-dimensional latent space (LatentMoE). Each token picks 16 of 512 routed experts plus one shared expert. The result: the KV cache is 1/24 of Hy3 and 1/6 of Hy4. On long contexts, memory and bandwidth are usually bounded by KV cache, so this is what decides whether a 1M window is affordable to actually serve.
Second, output tokens were halved. Tencent says reinforcement learning in post-training cut average end-to-end task output tokens by 50%. For an output-billed API, that matters more than a discount on the sticker price.
Third, decoding got fast. Combining megakernels, 4-bit quantisation and DFly2 speculative decoding, Tencent claims decoding above 150 tokens per second; TechNode cites a figure of about 200 tokens per second. Those are different sources with different measurement conditions, so do not treat them as the same claim.
How cheap, by Tencent's own accounting
Tencent did not lead with public leaderboards. It built an internal benchmark called WorkIndex, covering eight task sets: creating slides, editing slides, human-AI co-writing of slides, human-AI co-writing of Markdown, document editing, spreadsheet repair, audio and video processing, and visual understanding.
On that benchmark:
- Hy4 flash: 42.1
- Hy4 preview (770B total, 49B active): 39.7
- Hy4 (released November 4): 51.4
So it edges past the previous preview build but stays well below the full Hy4. On cost per task, Tencent's numbers are:
- 1/13 of Hy4 preview
- 1/20 of Hy4
- 1/8 of DeepSeek-V4.1-Flash-1017 (Max tier, peak-hour pricing)
A second set of numbers comes from a blind evaluation with 85 internal Tencent WorkBuddy users across 259 tasks in 27 categories. Hy4 flash scored 2.79 out of 4 on user satisfaction against 2.81 for Claude Haiku 5.5, at RMB 0.26 per task, with a median end-to-end time of 6.7 minutes — the lowest cost and shortest time among the models evaluated.
Three caveats apply to that study: the participants were Tencent employees, the tasks were WorkBuddy-style office work, and satisfaction was a subjective four-point score. A gap of 2.79 versus 2.81 is effectively a tie, not a precise 0.02-point deficit.
First-day numbers: the part that comes from media reports
The following is compiled from multiple media reports (TechNode, TokenPost, Writeble and others), not from the press release itself:
- TechNode, citing Tencent, says Hy4 flash processed more than 50 trillion tokens on its first day, the highest first-day usage of any model in the Hunyuan family.
- On OpenRouter's daily LLM leaderboard, the model ranked first on November 17.
- On WorkBuddy, first-day usage was 10 times that of Hy4 preview, and the model has helped complete more than 50 million real user tasks.
- TechNode dates the launch to November 16; the press release is datelined Shenzhen, November 17. A timezone difference explains the discrepancy, and both dates are circulating.
These are vendor-reported operating metrics with no third-party audit. Read them as a signal of adoption, not as evidence of capability.
How to use it: getting the weights and running them without tripping
Repository sizes, measured by summing the files listed on Hugging Face rather than taken from official documentation:
tencent/Hy4-flash: about 440GB across 200 filestencent/Hy4-flash-FP8: about 287GBtencent/Hy4-flash-base: about 439GB
Tencent recommends vLLM and SGLang, and ships prebuilt images for both:
- vLLM:
vllm/vllm-openai:hy4-flash, with the official example using FP8 weights,--tensor-parallel-size 8and--speculative-config '{"num_speculative_tokens":5,"method":"dfly"}' - SGLang:
lmsysorg/sglang:hy4-flash, published as a multi-architecture image for x86 and Arm
In other words, the reference configuration starts at eight GPUs with tensor parallelism. Single-GPU use depends on community quantisations: Hugging Face already hosts GGUF builds (the official AngelSlim GGUF repo has passed 50,000 downloads), plus MLX 4-bit and 8-bit conversions and NVFP4 variants.
A few practical pitfalls:
- transformers support has not landed on main yet. The model card says it is coming soon and points to a specific pull request branch for now.
- The base model is not a chat model. The card states plainly that
Hy4-flash-baseis pretrained, not instruction-tuned; the chat and reasoning examples apply only to Hy4 flash and the FP8 build. - The instruct models accept a
reasoning_effortsetting:"high"is the default with thinking on, and"no_think"gives direct answers. - Recommended sampling is temperature 0.9 and top_p 1.0.
Licence Section 10: the clause that will actually stop people
Hy4 flash uses the Tencent Hy License v1.0. The licence text says it is derived from Apache 2.0 with modifications, chiefly additional commercial terms for large model-as-a-service businesses in Section 10, and states explicitly that it is not the Apache License and is not endorsed by the Apache Software Foundation.
What Section 10 actually says:
- MaaS is defined as giving third parties remote access to model inference or fine-tuning, for example through an API that lets them submit inputs or specify parameters or training data. Two cases fall outside that definition: end-user products whose model capability is confined to specific features or workflows without general-purpose model access, and simply relaying requests to an independently operated third-party model service.
- A separate written licence is required in either of two cases: worldwide gross revenue above $50 million in any consecutive 12-month period, whether or not that revenue comes from the model; or use in a product or service with more than 100 million monthly active users in any calendar month.
- Internal use is exempt. Section 10.2 does not apply when the model, derivatives and outputs are all kept away from third parties.
In practice: an internal document or spreadsheet assistant is unaffected. A public API business large enough to trip those thresholds needs to negotiate first, at the address Tencent provides, [email protected]. The structure resembles the Llama community licence — stricter than MIT or Apache, but keyed to revenue and monthly actives rather than headcount.
Who it's for: use cases that fit
Tencent's own recommendation list is specific: deep search, document processing, slide creation, simple frontend demos, and computer use. The eight WorkIndex task sets point the same way — making and editing decks, co-writing documents, repairing spreadsheets, processing audio and video, understanding images.
A simple test: is the deliverable a thing you can hand over, or a project that runs for days? Weekly reports, contract redlines, merging a pile of spreadsheets into one, turning a meeting recording into minutes — these converge inside a single session, lean on long context, and repeat often, which is exactly where a 1M window and a low per-task cost pay off.
Who it's not for
- Long-horizon software engineering. Tencent itself points complex, long-running coding work at Hy4; Hy4 flash is positioned for everyday productivity.
- Anyone who needs publicly comparable benchmark scores. The headline benchmark is internal, published only as a summary image with no per-task public numbers, so nobody outside can reproduce it.
- Anyone hoping for a cheap single-machine trial. BF16 weights are 440GB and the reference setup is eight GPUs, so individual developers are pushed onto community GGUF and MLX builds that Tencent does not vouch for.
- Anyone who needs unrestricted commercial hosting. Section 10 rules out simply standing up a public API service at scale.
Cost and time estimates
A rough call at Tencent's list prices: 50,000 input tokens plus 5,000 output tokens costs about $0.005 (roughly $0.0035 in, $0.0014 out). That lines up with the RMB 0.26 per-task figure from the blind evaluation — office tasks tend to need several turns and tool calls.
Self-hosting shifts the cost from tokens to memory: FP8 weights across eight GPUs is the documented starting point, and rental cost has to be priced per hour on your own terms.
Next steps
- First, decide whether you need cheap high-frequency office work or long-horizon coding. The first is Hy4 flash territory; the second is Hy4.
- Second, do not start by pulling 440GB. Run the official FP8 image on eight GPUs against three of your own real tasks and judge the deliverables, not the 42.1 headline.
- Third, if you sell API access, check your revenue and monthly actives against Section 10 before deciding whether to start a licensing conversation.
- Fourth, if you plan to keep pretraining or fine-tuning on
Hy4-flash-base, remember it is not a chat model — chat templates and sampling settings apply to the instruct builds.
Sources
- Tencent press release (PR Newswire, November 17, 2026): https://www.prnewswire.com/news-releases/tencent-launches-hy4-flash-to-power-high-speed-cost-efficient-ai-deployment-for-everyday-workflows-302941875.html
- Hy4 flash model card and weights: https://huggingface.co/tencent/Hy4-flash
- Hy4 flash-FP8: https://huggingface.co/tencent/Hy4-flash-FP8
- Tencent Hy License v1.0: https://huggingface.co/tencent/Hy4-flash/blob/main/LICENSE
- Official code repository: https://github.com/Tencent-Hunyuan/Hy4-flash
- TechNode (November 19, 2026, weights release): https://technode.com/2026/11/19/tencent-releases-hy4-flash-model-weights-for-local-deployment/
- TechNode (November 19, 2026, 50 trillion first-day tokens): https://technode.com/2026/11/19/tencent-says-hunyuan-hy4-flash-surpasses-50-trillion-tokens-in-first-day-usage/
- TokenPost (November 18, 2026): https://www.tokenpost.com/news/technology/60137