跪拜 Guibai
← All articles
AI Coding · OpenAI · AIGC

Three Chinese Multimodal Models Go Head-to-Head on Real Production Tasks

By 猫猫头啊 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Multimodal models are crossing from demo to production, and the bottleneck is no longer accuracy — it's latency and cost per call. These results show that on structured extraction tasks, the cheapest model can be the fastest without sacrificing quality, which changes the default choice for high-volume agent and API workloads.

Summary

Three domestic Chinese multimodal models — Step 3.7 Flash, Qwen3.6-flash, and MiniMax M3 — were benchmarked on the same production tasks with identical prompts and parameters. The first task asked each model to reconstruct a 10-step WeChat Mini Program login flow from a single diagram. The second required extracting 12 structured JSON fields from an electronic invoice.

All three models produced correct outputs with no errors across both scenarios. Quality was indistinguishable. The differences emerged in speed and cost: Step 3.7 Flash was consistently faster and cheaper, with the lowest token consumption and API latency in every test. Invoice extraction cost less than one Chinese cent per call on Step 3.7 Flash.

The evaluation framework itself is worth noting. Rather than relying on public leaderboard scores, the tests measured three production-critical dimensions — one-shot output quality, end-to-end response time, and total cost including any manual follow-up. This mirrors the criteria any team would use when deciding whether a model is safe to wire into an agent loop or a customer-facing API.

Takeaways
Step 3.7 Flash, Qwen3.6-flash, and MiniMax M3 all extracted a 10-step login flow from a flowchart image with correct logic; Qwen merged two steps but stayed logically sound.
On invoice extraction, all three models returned perfectly accurate JSON for 12 fields with zero errors.
Step 3.7 Flash was the fastest in both scenarios: 15s vs 19–20s for the flowchart, and 5.6s vs 6.1–7.4s for the invoice.
Step 3.7 Flash consumed the fewest tokens and had the lowest cost — ¥0.0246 for the flowchart task and ¥0.0060 for the invoice, making structured extraction cheaper than one Chinese cent per call.
MiniMax M3 was the most expensive model in both tests, with token costs roughly 2.8× higher than Step 3.7 Flash on the flowchart task.
Quality stability was identical across all three models; no model required follow-up corrections or re-prompting.
Conclusions

Public benchmark scores are a poor proxy for production readiness. All three models scored well on leaderboards, but the real differentiator was cost and speed, not accuracy.

The evaluation criteria — one-shot quality, latency, and total cost including human intervention — are a practical template for any team vetting a model for agent or API use.

Multimodal models have reached a point where structured extraction from invoices and diagrams is a solved problem; the remaining competition is on price and throughput, not capability.

Concepts & terms
Multimodal model
An AI model that can process and reason over multiple types of input — typically text plus images, audio, or video — rather than text alone. In these tests, the models ingested screenshots and returned structured text.
Token consumption
The total number of tokens (word pieces) processed by a model call, including input, output, and cache reads/writes. It directly determines API cost, since providers charge per token.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗