Tencent Hy4 flash Benchmark Results for Data Analysis

Nov 16, 2026: Tencent's new low-cost Hy model, tested on real-world data analysis.

At Plotly, we keep a close eye on model releases. Our agentic analytics products, Plotly Studio Desktop and Plotly Studio Embedded, and our open source libraries Plotly and Dash work with any model, and our customers look to us for recommendations.

On November 16, Tencent released Hy4 flash, which its announcement (in Chinese) calls the most cost-effective Hy model yet. It has 224 billion parameters in total, 64 billion of them what Tencent calls Engram parameters, and 12 billion active. It's multimodal, with a context window of 1 million tokens. On OpenRouter, where Tencent is the only provider, it costs about $0.044 per million input tokens and $0.18 per million output tokens, about a tenth of Hy4's price per input token. On Tencent's own API it's ¥0.5 per million input tokens and ¥2 per million output tokens.

Tencent says that in pretraining it added hybrid attention, LatentMoE and Engram architecture features to Hy4 flash, and that in post-training it gained its general intelligence "through self-distillation of Hy4", then used reinforcement learning to cut the length of its end-to-end task output by 50%. Tencent recommends Hy4 flash for most everyday work, such as deep search, document processing, generating slides, simple front-end work and computer use, and Hy4 for long-horizon software engineering.

We ran it through our data analytics exam. Hy4 flash passed 39 of 43 questions (91%, an A) for $0.13, with a median of 29 seconds a question. That makes it the cheapest A-grade model we've tested: the previous cheapest, Particle-1, cost $0.35. Only GPT-6 Luna, at 38 for slightly less, is cheaper, so Hy4 flash joins the cost-accuracy frontier between GPT-6 Luna and Particle-1. Against Tencent's own Hy4, it gives up three questions for about a sixteenth of the cost, and it's about four times as fast.

Data analytics benchmark cost-accuracy frontier with every model labelled and Hy4 flash circled at 91% for $0.13

Hy4 flash vs. Hy4, GPT-6 Luna and Particle-1

Hy4 flash

Hy4

GPT-6 Luna

Particle-1

Score

39 / 43

42 / 43

38 / 43

40 / 43

Analytics questions

20 / 20

20 / 20

20 / 20

20 / 20

Ambiguous questions

5 / 5

5 / 5

5 / 5

5 / 5

Impossible requests

5 / 5

5 / 5

5 / 5

5 / 5

Data quality

3 / 3

3 / 3

3 / 3

2 / 3

Statistics traps

4 / 5

5 / 5

3 / 5

5 / 5

Hidden needles

2 / 5

4 / 5

2 / 5

3 / 5

Open weights

Announced

Announced

No

No

Cost of one full run

$0.13

$2.11

$0.13

$0.35

Per person per year*

$8

$123

$7

$21

Median time per question

29 seconds

117 seconds

18 seconds

16 seconds

*1,000 people each asking 10 questions a day, 250 working days a year, at OpenRouter's prices when each model ran.

Hy4 flash ties three other models at 39: Ember-1, Claude Fable 5.1 and Thinking Machines' Inkling 2 Small. That's the bottom of the A range, and 17 of the 51 models on our board score higher. None of the models at 39 or above costs less.

One run per model is a single sample. In our repeat runs of other models, scores moved by up to three points between runs, so treat a one- or two-point gap as close to a tie.

What it got right

Hy4 flash was perfect on four of the six kinds of question. It got all 20 straightforward analytics questions, it asked about or stated its assumptions on all five ambiguous questions, and it turned down all five impossible requests instead of inventing an answer. It also caught all three data quality problems we planted in the warehouse, as GPT-6 Luna did. Particle-1 caught two.

It got four of the five statistics traps, where the obvious calculation gives a misleading answer. That's one more than GPT-6 Luna and one fewer than Hy4 and Particle-1.

It's efficient, too. It made about 8 model calls per question and wrote about 4,800 output tokens per question, and 88% of the input tokens it read came from the provider's cache. The run had no model errors.

What it got wrong

It missed four questions, and we checked each one by hand against the transcript and the answer key. All four were genuine misses, not grading problems.

Three were hidden patterns, the hardest part of our exam. Each one points toward something unusual in the data without naming it, and the model has to go looking. In two of them, Hy4 flash never isolated the pattern we planted. In the third, it found the right place to look but stopped short of naming the cause its own query had shown. On these questions it ran about 33 queries each, above the median of about 26 on our board but about half as many as Hy4, which ran 64 and found four of the five.

The fourth was a statistics trap. Asked to pick out unusual values, it ranked the top ones instead of setting a threshold based on how the values were distributed.

Cost and speed

At $0.13 a run, Hy4 flash comes to about $8 per person per year at our rollout numbers, against $123 for Hy4. GPT-6 Luna costs about 3% less and scored one point lower. Particle-1 scored one point higher for about 2.7 times the cost.

Tencent compares Hy4 flash with Claude Haiku 5.5, DeepSeek V4.1 Flash and GLM 5.3 Flash. On our exam it beat all three, for less: Claude Haiku 5.5 scored 37 for $0.38, DeepSeek V4.1 Flash 36 for $0.46, and GLM 5.3 Flash 38 for $0.69.

Its median question took 29 seconds, and the whole exam about 33 minutes, against almost three hours for Hy4. That puts it in the middle of our board: 27 of the 51 models have a faster median, including GPT-6 Luna at 18 seconds and Particle-1 at 16.

If you're choosing between Tencent's two models for data analysis, Hy4 is the more accurate, and it found twice as many hidden patterns (more in our Hy4 post). Hy4 flash answers about four times as fast for a sixteenth of the cost, and it still scores an A.

Tencent's models over time

We've tested two Tencent models, shown here with every other model in grey. Hy4, released on November 4, scored 42 for $2.11, with a median of 117 seconds a question. Hy4 flash, twelve days later, trades three of those questions for about a sixteenth of the cost and about four times the speed.

Accuracy vs. cost of every Tencent model we've tested, with all other models in grey
Accuracy vs. median time per question for every Tencent model we've tested, with all other models in grey

The Test Harness

We run every model in Plotly Studio Embedded, a drop-in analytics agent that runs on Dash apps. This harness is tuned to run data analytics queries and make charts with multi-turn tool calling, with less emphasis on writing text. It's meant to put the data front and center.

Claude Sonnet 5.5 in Plotly Studio Embedded answering "Which is our best-performing wind farm?"

Frequently asked questions

How good is Hy4 flash at data analysis?

On PlotlyDataAnalyticsBench it passed 39 of 43 questions (91%, an A), tied with Ember-1, Claude Fable 5.1 and Inkling 2 Small. It was perfect on analytics questions, ambiguous questions, impossible requests and data quality problems, and it lost its points on three hidden patterns and one statistics trap.

Is Hy4 flash better than Hy4?

Not on accuracy. Hy4 scored 42 against 39 and found four hidden patterns against two. But Hy4 flash's run cost $0.13 against $2.11, and its median question took 29 seconds against 117.

How much does Hy4 flash cost?

On OpenRouter, about $0.044 per million input tokens and $0.18 per million output tokens, with a 1 million-token context window. Our whole exam cost $0.13, or about $8 per person per year at 10 questions a day. It's the cheapest model we've tested that scored an A.

Is Hy4 flash open source?

Not yet. Tencent says it plans to open-source Hy4 flash, but it hasn't released the weights or said when it will. For now it's available through Tencent's API and OpenRouter. Tencent has also announced but not yet released open weights for Hy4 (more in our open vs. closed models post).

How does Hy4 flash compare with other low-cost models?

It's on the cost-accuracy frontier: GPT-6 Luna is the only model we've tested that costs less, and it scored 38. Claude Haiku 5.5, DeepSeek V4.1 Flash and GLM 5.3 Flash all scored lower and cost more. GPT-6 Luna and Particle-1 are faster, at 18 and 16 seconds a question against 29.

About the benchmark

We run new models through the same exam as they're released, and we'll keep sharing what we find. We ran Hy4 flash through OpenRouter on the day it arrived there, with the same harness and grading as every other model.

To see all 51 models on one chart and in a full results table, along with cost, speed and how the exam works, see Introducing DataAnalyticsBench.

The exam is private and we're sharing it with researchers. We're keeping it private so that the answer key doesn't make its way into the training set. For the same reason, the data is fabricated: the warehouse is a synthetic wind-energy fleet, built to look like a real operator's data, so its answers don't exist anywhere else.

If you'd like access, reach out to Chris Parmer at Plotly.

bluesly
x
instagram
youtube
medium
fcaebook

Product

© 2026
Plotly. All rights reserved.
Cookie Preferences
aicpa/soc2