TabbitBlog

HY4 Flash vs HY4: Price, Benchmark Results, and Fit

Compare Tencent's two Hy4 models with same-provider rates, Plotly's same-harness data test, and Tencent's workload guidance.

In this article
  1. Key takeaways
  2. HY4 Flash vs HY4 at a glance
  3. What Plotly's same-harness test found
  4. Choose by workload, not model size
  5. Price per token is not price per completed job
  6. Verdict

HY4 Flash is much cheaper per token on Tencent TokenHub; HY4 scored higher in Plotly's one-run data-analysis comparison. That is the most useful direct answer today. TokenHub list rates make Flash 12x lower on uncached input, 9x lower on output, and 15x lower on cache-hit input. Plotly found Flash three points behind HY4, but the one run cost about one-sixteenth as much and took about one-quarter of the median time per question. Neither result predicts every workload.

Tencent recommends Flash for routine everyday work and HY4 for long-horizon software engineering. Start with that split, then verify against your own task. For a deeper single-model view, read the HY4 Flash review; the pricing guide separates list rates from the current promotions.

Key takeaways

  • Flash has lower TokenHub unit rates: ¥0.50/¥2/¥0.02 per million input/output/cache-hit tokens versus HY4's ¥6/¥18/¥0.30.

  • Both models list 1M context; TokenHub specifies 960k max input and 64k max output for each.

  • Plotly scored HY4 42/43 and Flash 39/43 in one private test.

  • The reported cost and latency gaps belong to that test setup, not a general price or speed guarantee.

  • Tencent's open-weight plans are not the same as released weights and license.

HY4 Flash vs HY4 at a glance

DimensionHY4 FlashHY4
Total / active parameters224B total, 12B active770B total, 49B active
TokenHub context limits1M context; 960k max input / 64k max output1M context; 960k max input / 64k max output
TokenHub list rate per 1M¥0.50 input / ¥2 output / ¥0.02 cache hit¥6 input / ¥18 output / ¥0.30 cache hit
Plotly one-run result39/43; $0.13; 29 sec median/question42/43; $2.11; 117 sec median/question
Tencent's stated fitEveryday office work and frequent callsLong-horizon software engineering

Rates are from Tencent's TokenHub tariff. Flash's 50% promotion ends Nov 30, 2026 at 12:00 Beijing time; HY4's separate 50% promotion runs through Dec 31, 2026 at 23:59:59 Beijing time. Do not compare their temporary offers without recording those different deadlines.

What Plotly's same-harness test found

Plotly ran each model once through the same analytics agent and scoring rules. HY4 Flash earned 39/43; HY4 earned 42/43. Flash's reported run cost $0.13 versus $2.11 and took 29 seconds per question at the median versus 117 seconds. Flash missed three hidden-pattern questions and one statistical trap; HY4 found four of five hidden patterns.

The exam uses private synthetic wind-farm data, so readers cannot rerun the exact questions. Plotly also says results from repeated tests of other models can move by up to three points. The 3-point difference here should be read as an observed result, not a universal quality gap. See the full Plotly DataAnalyticsBench report.

Choose by workload, not model size

Your workloadSensible first candidateWhy / what to verify
Repeated extraction, summaries, routine calculationsHY4 FlashLower unit rates; validate against fixed expected output
Long coding task with many tool turnsHY4Tencent recommends it for long-horizon software engineering; test the full repository task
Finding hidden anomalies in dataNeither without reviewPlotly's one run missed three hidden patterns on Flash and one on HY4
Budget-constrained, high-volume agent requestsHY4 Flash trialMeasure tokens, cache hits, retries, and success rate
Local deploymentWait for published artifacts or choose another available routeTencent's announcement says Flash weights are planned, not currently downloadable

One Reddit commenter said 12B active parameters looked attractive for local use “if the weights drop.” That is conditional interest, not evidence of a current local install. Another reply stressed that open weights were not yet released. Read the launch thread.

Price per token is not price per completed job

At list rates, a million uncached input tokens cost ¥0.50 with Flash and ¥6 with HY4. But a job's bill also depends on output volume, cache hits, call count, and retries. Plotly's specific benchmark used about eight calls and 4,800 output tokens per question, and 88% of input came from cache. Those conditions shape the reported total.

The reported benchmark costs were $0.13 and $2.11, in OpenRouter, not Tencent TokenHub's CNY rate card. Use the same provider and real usage records when comparing a workload. The HY4 Flash vs DeepSeek V4.1 Flash comparison discusses provider-matched price comparisons.

Verdict

Start with HY4 Flash for bounded, repetitive work where you can check results and need lower token rates. Use HY4 as the candidate for long-horizon engineering, following Tencent's guidance, and verify the end-to-end result. In analytical work, add a second check for hidden patterns regardless of model.

For model access and setup, see HY4 Flash resources and Tencent Hy4 in a signed-in browser. For other client paths, see DeepSeek V4.1 Flash, its browser workflow, the DeepSeek Harness plugin guide, the Feishu Base example, Tabbit CLI setup, and the browser-agent first-task checklist.

Tabbit Browser offers Hy4 Flash. Download Tabbit, then select Hy4 Flash from the model list in Chat to start using it.

Tabbit Browser

FAQ

Is HY4 Flash cheaper than HY4?

At Tencent TokenHub list rates, HY4 Flash costs ¥0.50/M input, ¥2/M output, and ¥0.02/M cache hit; HY4 costs ¥6/M, ¥18/M, and ¥0.30/M. Those are token-price ratios, not task-cost guarantees.

Which model scored higher in Plotly's test?

HY4 scored 42/43 and HY4 Flash 39/43 in Plotly's one-run private DataAnalyticsBench. HY4 Flash cost $0.13 in that test versus $2.11 for HY4, and its median time per question was 29 versus 117 seconds.

Does HY4 Flash have the same context size as HY4?

Tencent lists both at up to 1M context. On TokenHub, both have a 960k maximum input and 64k maximum output; other providers can expose different limits.

Should I use HY4 or HY4 Flash for coding?

Tencent recommends HY4 for long-horizon software engineering and HY4 Flash for routine everyday work. Treat that as vendor guidance and compare them on your own codebase and acceptance tests.

Can I run HY4 Flash locally?

Tencent announced plans to open-source HY4 Flash but did not give a release date in the November 16 announcement. Confirm weights and license are available before planning local inference.

Take the next step

Let Tabbit work alongside you.

Research across tabs, automate repetitive browser work, and keep every piece of context within reach.