Tencent Hy4 Flash: Benchmarks, Cost per Task and WorkIndex
Tencent's Hy4 flash scores 42.1 on WorkIndex at 1/13 the cost of Hy4 preview. Benchmarks, blind test results and architecture explained.
For Engineering and Executives

How does Tencent Hy4 flash perform, and what does it cost per task?
Hy4 flash is Tencent's most cost-effective Hy model so far. On Tencent's internal WorkIndex benchmark it scores 42.1, slightly above Hy4 preview (39.7), while its average cost per task is about 1/13 of Hy4 preview and 1/20 of the full Hy4. In a blind test with Tencent's own employees, it scored 2.79 out of 4 on user satisfaction, against 2.81 for Claude Haiku 5.5, at RMB 0.26 per task. All of these numbers come from Tencent's model card, so treat them as the vendor's own results.
TL;DR
- Hy4 flash scores 42.1 on WorkIndex, ahead of Hy4 preview (39.7) but behind the full Hy4 (51.4).
- Its average cost per task is 1/13 of Hy4 preview, 1/20 of Hy4, and 1/8 of DeepSeek-V4.1-Flash-1017 (Max tier, peak-hour pricing).
- In a blind test, users rated it 2.79/4 versus 2.81/4 for Claude Haiku 5.5, with the lowest cost (RMB 0.26 per task) and shortest median completion time (6.7 minutes) among the models tested.
- On base-model academic benchmarks it beats DeepSeek-V4-flash-Base on several tests and loses on others, so the picture is mixed.
- It is a mixture-of-experts model with 220B total parameters but only about 12B active per token, plus a 1M-token context window and text and image input.
- Tencent recommends it for everyday work like document processing, slides and deep search, and the full Hy4 for hard, long software engineering jobs.
- The weights are open, so you can download and run them yourself.
What is Hy4 flash?
Hy4 flash is a language model from Tencent's Hy family. It accepts text and images, produces text, and supports a context window of up to 1 million tokens. (A token is a small chunk of text, roughly a short word or part of one. The context window is how much the model can read at once.)
The headline design idea is "big knowledge, small bill." The model has 220B total parameters, but 64B of those belong to a special memory component called Engram. Only about 12B parameters are active for any given token. Fewer active parameters per token means less computing per word generated, which is where much of the cost saving comes from.
Tencent released three versions as open weights on Hugging Face, ModelScope and GitCode: the instruct model (Hy4 flash), an FP8 quantized version, and the pre-trained base model. The license is the Tencent Hy License, so read it before using the model commercially.
How does Hy4 flash compare with Hy4 and Hy4 preview on cost?
This is the part of the card that matters most to a budget holder. Tencent measured average cost per task on WorkIndex across its models.
| Model | WorkIndex score | Cost per task relative to Hy4 flash |
|---|---|---|
| Hy4 flash | 42.1 | 1x |
| Hy4 preview | 39.7 | about 13x |
| Hy4 | 51.4 | about 20x |
| DeepSeek-V4.1-Flash-1017 (Max tier, peak-hour pricing) | not stated in the text | about 8x |
The multiples are simply the inverse of Tencent's stated ratios. The card gives the DeepSeek comparison only as a cost ratio, and specifies the Max tier at peak-hour pricing, so it is not a general statement about DeepSeek's cost.
The trade is clear. Hy4 flash gives up 9.3 points of WorkIndex to the full Hy4 and cuts cost per task by a factor of 20. Against Hy4 preview it is both better and far cheaper. If your tasks are ordinary office work, that gap may be worth paying. If your tasks are hard, it may not be.
What is WorkIndex, and how far should you trust it?
WorkIndex is an internal benchmark Tencent built, grounded in typical tasks from WorkBuddy (Tencent's work-assistant product, per the card's wording). It has 8 evaluation sets:
- Slide creation
- Slide editing
- Human-AI slide co-writing
- Markdown human-AI co-writing
- Document editing
- Spreadsheet repair
- Audio and video processing
- Visual understanding
Tencent says it focused on deliverable quality, single-turn output length and task turnaround time, rather than public benchmark scores "detached from everyday work."
That framing is reasonable. Most office work looks like these tasks, not like a math olympiad. But the caveat is just as plain: Tencent designed the test, chose the tasks and ran it. Nobody outside the company has reproduced WorkIndex, and the card does not publish the tasks. A good score tells you the model suits the work Tencent cares about. It does not guarantee the same result on your documents.
The card also reports an "Open Benchmarks" group of 7 public agent tests (DeepSWE, ProgramBench, Terminal-Bench 4.0, APEX-Agents, AutomationBench, HLE and OSWorld 2.1), plotted against cost. The numbers sit in charts rather than text, so we do not quote them here.
What did the blind test show?
The blind evaluation used 85 internal WorkBuddy users and 259 tasks across 27 categories. Users rated outputs without knowing which model produced them.
| Measure | Hy4 flash | Claude Haiku 5.5 |
|---|---|---|
| User satisfaction (out of 4) | 2.79 | 2.81 |
| Median end-to-end completion time | 6.7 minutes | not stated in text |
| Cost per task | RMB 0.26 | not stated in text |
Tencent says Hy4 flash recorded the lowest cost per task and the shortest median completion time among the evaluated models. A 0.02 gap in satisfaction is small, and the card does not claim it is statistically meaningful in either direction.
Keep the limits in mind. The raters were Tencent employees, the tasks were Tencent's, and the evaluated set of models is not fully listed in the text. This is useful evidence that the model is competitive at low cost, not a neutral industry ranking.
How do the base-model benchmarks compare with DeepSeek-V4?
The card also compares the pre-trained base models (before instruction tuning) on few-shot tests. Here is a selection of Hy4 flash Base against DeepSeek-V4-flash-Base and DeepSeek-V4-Pro-Base.
| Benchmark | Hy4 flash Base | DeepSeek-V4-flash-Base | DeepSeek-V4-Pro-Base |
|---|---|---|---|
| MMLU | 88.7 | 89.7 | 91.1 |
| MMLU-Pro | 71.1 | 68.6 | 73.4 |
| SuperGPQA | 57.8 | 52.5 | 54.8 |
| SimpleQA | 37.5 | 31.3 | 52.1 |
| U-MATH | 69.8 | 57.4 | 64.8 |
| ARC-Challenge | 95.0 | 97.3 | 97.0 |
| MBPP-plus (code) | 73.7 | 76.8 | 79.1 |
The pattern: against the similarly named flash model, Hy4 flash wins on MMLU-Pro, SuperGPQA, SimpleQA and U-MATH, and loses on MMLU, ARC-Challenge and the coding test MBPP-plus. Against the larger Pro base model, it still wins on SuperGPQA and U-MATH but trails on most others, including a wide gap on SimpleQA, a factual recall test.
Tencent adds a disclaimer worth repeating. Its base model saw a lot of synthetic instruction data during mid-training, so it sometimes fails to follow the expected answer format in few-shot tests, which may distort scores. Also, bold marks in the card's table show the highest number in each row, not statistical significance. The full Hy4 base scores higher than Hy4 flash on nearly every row, as you would expect.
How is Hy4 flash built to be cheap?
The card describes savings at every stage. These are the main pieces.
Hybrid attention. Attention is how a model looks back over what it has already read. Hy4 flash mixes sparse attention (which picks up to 2,048 relevant tokens to look at) with sliding-window attention (which looks at only the last 512 tokens) at a 1:3 ratio. Most layers use the cheap local view, and a minority do the long-range lookup.
LatentMoE. In a mixture-of-experts model, each token is routed to a few specialist sub-networks instead of the whole model. Here, each token activates 16 of 512 routed experts plus one shared expert, and the experts work in a lower-dimensional space to save compute.
Engram. This adds learned representations of common two-word and three-word sequences, retrieved by hash lookup. It accounts for most of the 64B Engram parameters, and it adds stored knowledge without adding much per-token computing.
Tencent says these changes cut the KV cache (the memory a model uses to remember the conversation so far) to 1/24 of Hy3 and 1/6 of Hy4. Smaller memory use is what makes long contexts affordable.
On the training side, reinforcement learning cut average end-to-end output tokens by 50%. Because you pay per token generated, shorter answers directly lower cost. For serving, Tencent combines megakernels, 4-bit quantization and a built-in 0.85B-parameter drafter called DFly2 to push decoding speed past 150 tokens per second.
What does this mean for you?
If you use AI for everyday office work: the practical takeaway is not "switch to Hy4 flash." It is that cheaper models are now close enough on routine tasks that cost per task is a fair thing to ask about. Tencent recommends Hy4 flash for deep search, document processing, slide creation, simple frontend demos and computer use, and the full Hy4 for long-horizon software engineering.
If you manage a team or budget: cost per task, not price per token, is the number to track. A model that writes shorter answers and finishes sooner can be cheaper even if its listed price looks similar.
If you are technical or work with IT: the open weights mean you can test it on your own data. Tencent's card covers running it with Transformers, vLLM and SGLang, and a GGUF version is linked. Check the license terms first.
What to do this week: pick 20 real tasks from your own work, run them through two models, and record quality, time and cost for each. Vendor benchmarks narrow the shortlist. Your own tasks decide it.
Can you build your own model comparison tracker?
That last step is where most people reach for a spreadsheet, and it gets messy fast. A small internal app could do it better: a form where colleagues submit a task, a side-by-side view of two model outputs, a 1 to 4 rating, and a running total of cost and time per task. You can describe that to Remy in plain English and get a deployed app with a database and sign-in. Remy gives apps access to nearly 400 AI models without managing API keys, and AI usage is billed at cost, so you can run your own blind test with your own people.
Frequently Asked Questions
Is Hy4 flash better than Hy4?
No. On WorkIndex, Hy4 scores 51.4 and Hy4 flash scores 42.1. Hy4 flash is much cheaper, at about 1/20 the cost per task, and Tencent recommends the full Hy4 for complex work such as long-horizon software engineering.
How much does Hy4 flash cost per task?
In Tencent's blind evaluation, the cost was RMB 0.26 per task. On WorkIndex, the card gives cost only as ratios: 1/13 of Hy4 preview, 1/20 of Hy4, and 1/8 of DeepSeek-V4.1-Flash-1017 (Max tier, peak-hour pricing). Your own cost will depend on where and how you run it.
Can I run Hy4 flash myself?
The weights are released openly on Hugging Face, ModelScope and GitCode, including an FP8 version and a base model. The card describes deployment with Transformers, vLLM and SGLang, and the license is the Tencent Hy License. The card text we reviewed does not list hardware requirements, so check the official repository before planning.
Can I trust the benchmark results?
They are useful but come from Tencent. WorkIndex is an internal benchmark, and the blind test used Tencent's own employees and tasks. The base-model results use public benchmarks, but the card warns that format-following problems may distort some few-shot scores. Test the model on your own tasks before committing.
Useful for
How we source, write, and correct our articles. Editorial standards
Related articles
Live Nerf vs Margin Lab: Do AI Models Really Get Worse?

Why Is Nvidia So Dominant in AI Chips? The CUDA Moat Explained

Do AI Models Get Nerfed After Launch? What Labs Admitted
