
Tencent has released the weights of Hy4 flash, the most cost-effective model in its Hy (Hunyuan) family so far. It takes text and image input, handles up to 1M tokens of context and is built for everyday office work: deep search, document processing, slide creation, simple web prototypes and computer use.
In a blind evaluation that Tencent ran with 85 internal WorkBuddy users, Hy4 flash scored 2.79 out of 4 for user satisfaction – almost level with Claude Haiku 5.5 (2.81). Independent measurements are tougher: on the Vals AI composite Vals Index, the model scores 50.8%, 0.5 points below the open-weight DeepSeek-V4.1-Flash and 7.9 points below the closed Claude Haiku 5.5. Tencent is not chasing records here – it is betting on cost per task and speed.
What’s new in Hy4 flash
First, open weights. Three versions are out on Hugging Face, ModelScope and GitCode: the instruct model Hy4 flash, Hy4 flash-FP8 with weights quantized to 8 bits, and the base model Hy4 flash-base, which has not been instruction-tuned. The license is Tencent’s own Tencent Hy License, and the weights come with a complete fine-tuning pipeline.
Second, cost per task. On WorkIndex, Tencent’s internal benchmark of eight office task sets ranging from slides to spreadsheets and video, Hy4 flash scores 42.1, against 39.7 for Hy4 preview and 51.4 for the full Hy4. By Tencent’s calculation, an average task costs 13 times less than on Hy4 preview, 20 times less than on Hy4 and 8 times less than on DeepSeek-V4.1-Flash in Max mode at peak-hour pricing.
Third, short, fast answers. The model’s knowledge was transferred from Hy4 itself (self-distillation), and according to Tencent, reinforcement learning halved the average number of output tokens. The company says a combination of 4-bit quantization, megakernels and the built-in DFly2 draft model pushes decoding speed past 150 tokens per second. Reasoning mode is on by default and can be switched off with the reasoning_effort parameter.
Architecture: 12B active parameters out of 220B
Hy4 flash is a mixture-of-experts (MoE) model. The backbone accounts for 155.5B parameters, another 63.5B sit in Engram tables with learned representations of token bigrams and trigrams, 0.43B belong to a SigLIP2-based vision encoder and 0.85B to the DFly2 draft model. Each of the 41 MoE layers has 512 experts, and a token activates only 16 of them plus one shared expert.
Hybrid attention keeps long context cheap. Of the 42 layers, 12 use sparse attention that picks up to 2,048 of the most relevant tokens from the whole context, while the other 30 only see a window of the last 512 tokens. According to Tencent, the KV cache – the memory that holds the context – is 6 times smaller than Hy4’s and 24 times smaller than Hy3’s. DFly2 reuses this cache, so the speed-up needs no extra memory and does not get more expensive as the context grows.
Which models Hy4 flash competes with
The closest rivals are budget models with a context window of about 1M tokens: the open-weight DeepSeek-V4.1-Flash, which Tencent itself compares Hy4 flash against, and the closed Claude Haiku 5.5. The full Hy4 is in the table for scale. Hy4 flash had no Artificial Analysis score at the time of writing, while Hy4 scores 34.6 on the Artificial Analysis Intelligence Index – the same as Claude Haiku 5.5 at Xhigh reasoning effort. Vals AI measurements are the common denominator for all four models.
| Model | Vals Index, % | Terminal-Bench 4.0, % | Price, $ per 1M tokens (input / output) | 1,000 tasks, $ |
|---|---|---|---|---|
| Hy4 flash | 50.8 | 15.2 | 0.07 / 0.281 | 9.81 |
| DeepSeek-V4.1-Flash | 51.3 | 19.7 | 0.051 / 0.567 | 10.77 |
| Claude Haiku 5.5 | 58.7 | 36.4 | 0.10 / 0.50 | 15.00 |
| Hy4 | 59.4 | 39.4 | 0.834 / 2.501 | 108.41 |
Vals Index and Terminal-Bench 4.0 are Vals AI measurements; prices are from OpenRouter, for Tencent models without the current 50% discount. The cost of 1,000 tasks is our editorial estimate: 100K input and 10K output tokens per task, no caching, and the same token usage for every model.
Hy4 flash trails DeepSeek-V4.1-Flash and Hy4 on every benchmark they share, and Claude Haiku 5.5 on nine out of ten. The gap is widest in agentic work in the terminal: on Terminal-Bench 4.0, Haiku 5.5 and Hy4 score 2.4–2.6 times higher. On Vibe Code Bench v1.1 the gap is smaller: 74.8% against 84.7% for DeepSeek-V4.1-Flash and 90.4% for Haiku 5.5. Hy4 flash comes closest to DeepSeek-V4.1-Flash: the Vals Index difference is under one point, and by our estimate 1,000 tasks cost 9% less.
Tencent itself recommends Hy4 flash for everyday office tasks and the full Hy4 for long-horizon software engineering. The Vals AI numbers agree: Claude Haiku 5.5 or Hy4 are the better fit for agentic coding. Output length is also limited: up to 65,536 tokens on OpenRouter, against 128,000 for Haiku 5.5.
The strengths are price, size and open weights. You can deploy and fine-tune the model on your own hardware, and in total parameters it is 2.5 times smaller than DeepSeek-V4.1-Flash: 220B against 552B. The vLLM and SGLang examples in the model card use the FP8 version on 8 GPUs: the weights alone take roughly 220 GB (editorial estimate: one byte per parameter), so a single consumer graphics card won’t do. A compressed GGUF version is available for testing. For multilingual tasks, a reference point is Belebele, which covers German, Korean, Vietnamese, Spanish and Russian: according to Tencent, the base Hy4 flash scores 93.8% against 94.4% for DeepSeek-V4-flash-Base.
Pricing, speed and deployment
In the API, the model is available through OpenRouter, where it is served by a single provider, Tencent Cloud. The Tencent prices in the table are list prices: at the time of writing, a 50% discount with no stated end date is in effect, and cache reads cost $0.003 per 1M tokens. Function calling and structured JSON output are supported.
OpenRouter measures the speed at about 70 tokens per second with 2.85 s latency – half the 150 tokens per second Tencent claims. For Claude Haiku 5.5, the main providers deliver 148–185 tokens per second.
By our estimate, Hy4 flash is 11 times cheaper than Hy4 and 1.5 times cheaper than Claude Haiku 5.5, and with the discount 1,000 tasks would cost about $4.90. Hy4 flash tokens are 9–12 times cheaper than Hy4 tokens; the rest of the gap to the 20x that Tencent claims may come from shorter answers. For self-hosting there are prebuilt vLLM and SGLang images and a GitHub repository, while Transformers support has to be installed from a separate branch for now.
Precedent
Hy4 flash continues Tencent’s Hy4 generation: Hy4 preview came first, and less than two weeks before flash the company released the full Hy4, the first in the line with image understanding. DeepSeek tried the “flagship plus a cheap open-weight flash version” approach earlier: DeepSeek-V4.1-Flash came out about two months before Hy4 flash, and at the time of writing it is served by 28 providers on OpenRouter, with input prices starting at $0.051 per 1M tokens. If the pattern repeats, Hy4 flash will also get third-party providers and price competition (a forecast based on precedent).









