APIYI: Seed 2.0 Lite and Pro Benchmark Table, Vision Tiers, and Production Scenarios

One-sentence takeaway

APIYI's comparison table shows Lite approaching Pro on AIME, knowledge, vision, video, and coding tasks, and publicly documents low/high/xhigh vision tiers; these are benchmark figures compiled from a provider page and should be rechecked on the same API.

Use cases

• Suitable tasks: Document, receipt, and contract extraction; chart analysis; video understanding; search and recommendation; tool calling; and high-QPS production routing.

• Unsuitable tasks: Treating provider-page benchmarks as an independent rerun, or omitting human spot checks in financial or high-risk tasks.

• Applicable model version: APIYI's seed-2-0-lite-260228; the page also lists Seed2.0 Pro and Seed1.8 for comparison.

• Applicable client, agent, or API: APIYI OpenAI-compatible endpoint https://api.apiyi.com/v1; this is not a direct Volcengine Ark connection.

• Recommended reasoning tier and parameters: Vision inputs can use the low/high/xhigh tiers; the page does not disclose complete temperature, reasoning, or tool parameters.

Test environment

• Test provider: APIYI's model introduction and benchmark summary pages.

• Comparison models: Seed 2.0 Lite, Seed 2.0 Pro, and Seed 1.8.

• Test tasks: AIME 2025, MMLU-Pro, SWE-Bench Verified, LiveCodeBench v6, MathVision, MathVista, VideoMME, and COLLIE.

• Complete inputs/configuration: The page does not disclose the original prompts, sample size, number of sampling runs, tools, or evaluation scripts.

Inputs/configuration

• Input modalities: Text, images, and video; output type is text.

• Vision tiers: low is suitable for simple recognition and fast classification; high is the default tier for standard document and chart analysis; xhigh targets dense text, complex charts, and detail-rich scenarios.

• API: OpenAI-compatible; the model name listed on the page is seed-2-0-lite-260228.

• Knowledge cutoff: The page lists 2024-01; this field should be verified against the current provider's actual behavior and documentation.

Results

Values listed directly in APIYI's page table:

Based on these figures, the page positions Lite as costing about one-fifth as much as Pro, suitable for everyday production tasks and high-concurrency coverage, and presents MMLU-Pro as an example of a knowledge task where Lite scores higher than Pro.

Conclusion

This material supports Lite as a cost-saving alternative candidate to Pro for structured knowledge, visual understanding, and high-throughput multimodal scenarios; SWE-Bench and LiveCodeBench remain below Pro, so complex software engineering should retain an upgrade path. Vision tiers can be directly converted into cost/quality A/B variables.

Limitations

• APIYI is a third-party provider, and the page provides no original benchmark inputs, evaluation code, or evidence of independent verification.

• “About one-fifth the cost,” the knowledge cutoff, model name, and endpoint all reflect the provider page's wording and are not equivalent to Ark's current pricing or snapshot.

• The benchmarks in the table may come from the official Model Card; they cannot be labeled as an independent controlled test by APIYI.

Reproduction steps

1. Fix seed-2-0-lite-260228, the vision tier, temperature, max output, and request mode in APIYI.

2. Save the original inputs and outputs for AIME/MMLU/coding/vision/video tasks, and use the same prompts, tools, and timeouts with Pro.

3. Separately measure accuracy, field/code tests, video event recognition, tokens, latency, errors, and retry costs.

4. Run an A/B test with the same inputs across the high and xhigh vision tiers to determine whether the quality gain offsets the cost.

Source excerpts or observations (for compliant short quotations only)

• The page splits Lite's vision input into low/high/xhigh tiers rather than giving it a single “vision capability” label.

• The page lists Lite's and Pro's coding and knowledge scores together, making them suitable for a routing hypothesis, but they should not replace independent retesting.