Which Chinese Multimodal LLM Is the Most Efficient in Real‑World Use?

Evaluation Methodology

Three multimodal models (Step 3.7 Flash, MiniMax M3, Qwen 3.6‑flash) are tested on identical tasks with the same prompt, parameters and tooling; only the model varies. For each run three dimensions are recorded: quality (usable answer without follow‑up), speed (end‑to‑end latency), and cost (model price × token consumption).

Scenario 1 – Reconstruct Business Logic from a Flowchart

Input: screenshot of a WeChat mini‑program login flow (10 steps). Prompt: @微信小程序登录方案.png 获取图片中流程逻辑 Outputs:

Step 3.7 Flash – correctly identifies all 10 steps, matching the original diagram.

MiniMax M3 – also returns 10 correct steps.

Qwen 3.6‑flash – returns 9 steps, merging steps 3 and 4; overall logic remains correct.

Performance metrics:

API time: 15 s (Flash), 20 s (MiniMax M3), 19 s (Qwen 3.6‑flash).

Token consumption (Input / Output / Cache‑Read / Cache‑Write):

Flash – 728 / 1.1k / 54.4k / 0

MiniMax M3 – 27.9k / 1.2k / 228 / 0

Qwen 3.6‑flash – 251 / 1.9k / 0 / 28.6k

Token price (¥ per request): 0.0246 (Flash), 0.0688 (MiniMax M3), 0.0483 (Qwen 3.6‑flash).

All three models produce correct quality; Flash is faster and cheaper.

Scenario 2 – Structured Extraction from an Invoice

Task: extract key fields from an electronic invoice image and return a JSON object. Prompt:

请从这张票据图片中提取结构化信息，按照如下 JSON 结构返回：{ 发票类型:string, 发票号码:string, 开票日期:string, 发票金额:string, 税率:string, 税额:string, 项目名称:string, 购买方纳税人识别号:string, 购买方开户行:string, 销售方名称:string, 销售方纳税人识别号:string, 销售方开户行:string }

Results:

Step 3.7 Flash – correct extraction, 5.6 s, 1,409 tokens.

MiniMax M3 – correct extraction, 6.1 s, 2,216 tokens.

Qwen 3.6‑flash – correct extraction, 7.38 s, 2,008 tokens.

Metrics:

API time: 5.6 s (Flash), 6.1 s (MiniMax M3), 7.38 s (Qwen 3.6‑flash).

Token consumption (Input / Output / Cache‑Read / Cache‑Write):

Flash – 802 / 607 / 0 / 0

MiniMax M3 – 1,686 / 530 / 1,672 / 0

Qwen 3.6‑flash – 1,165 / 843 / 0 / 0

Token price (¥): 0.0060 (Flash), 0.0086 (MiniMax M3), 0.0075 (Qwen 3.6‑flash).

Again, quality is identical; Flash shows the lowest latency and token cost.

Aggregated Comparison

Across the two scenarios the models are ranked on the three dimensions:

Speed – Fast (Flash), Medium (MiniMax M3, Qwen 3.6‑flash).

Token consumption – Low (Flash), High (MiniMax M3), Medium (Qwen 3.6‑flash).

Cost per request (¥) – Cheap (Flash), Expensive (MiniMax M3), Medium (Qwen 3.6‑flash).

Stability – Excellent for all three models.

Thus, while all models meet the quality requirement for production use, Step 3.7 Flash consistently offers better speed and lower token cost, making it the preferred choice for high‑frequency agent or API deployments. Users should validate with their own data before final integration.