Qwen3.5-Plus: Qubrid's Same-Image, Same-Prompt Latency and Token Comparison

One-sentence takeaway

In Qubrid's single-run Playground vision test, Qwen3.5-Plus had a faster time to first token and total response, but spent most completion tokens on reasoning and produced shorter output text; this does not establish quality for all tasks.

Use cases

• Suitable tasks: Initial screening where visual input, low latency, and raw generation speed matter; also reproducing a baseline with the same image, prompt, and provider.

• Unsuitable tasks: Using one image and one run to prove complex Agent, long-context, or coding ability; or comparing latency and tokens across providers.

• Applicable model versions: Qwen/Qwen3.5-Plus in Qubrid Playground; the specific backend snapshot was not disclosed.

• Applicable clients, Agents, or APIs: Qubrid AI Playground; the article also provides an example for Qubrid's OpenAI-compatible interface.

• Recommended reasoning level and parameters: For Qwen3.5-Plus, the article recommends temperature=0.6 and top_p=0.95; this is the article's provider configuration recommendation, not a recommendation from Alibaba Cloud's official model page.

Test environment

• Platform: Qubrid AI Playground, Vision mode.

• Input: The same image of “an origami boat on a blue-gray surface”; the user prompt was Describe what you see in this image..

• Configuration: Model Reasoning was enabled for both models; the article records prompt tokens, completion tokens, reasoning tokens, output-text tokens, TTFT, total duration, and tokens/s.

• Comparison: Qwen 3.5 Plus and Qwen 3.6 Plus; this article focuses on the original observations for Qwen3.5-Plus.

Input/configuration

Visual input: the same origami boat image
User prompt: Describe what you see in this image.
Model Reasoning: True
Model: Qwen/Qwen3.5-Plus

Results

Calculated from the article's data, reasoning tokens account for approximately 91.3% of Qwen3.5-Plus's completion tokens (1,858/2,036), and output text accounts for approximately 8.7%; this is the allocation in that run, not a fixed model ratio.

Conclusion

This case supports a narrow conclusion: in Qubrid's single-run, same-image, same-prompt test, Qwen3.5-Plus had lower total latency and higher generation speed than the comparison model, but its answer text was shorter and its reasoning share was higher. Production Agent evaluation should continue with tool-call success, structured output, repeatability, and long-context retrieval.

Limitations

• The article does not disclose the image file, random seed, complete backend snapshot, concurrency, hardware, or number of repetitions; strictly speaking, only an approximate experiment can be reproduced from the public steps.

• Timing and token statistics from the Qubrid provider cannot directly represent Alibaba Cloud Bailian's native endpoint.

• The article's table extends Qwen3.5-Plus's multimodal description to audio, while Alibaba Cloud's official model page lists text, image, and video as inputs; this conflict should be resolved in favor of the official capability table, and audio capability should not be presented as verified fact.

Reproduction steps

1. Select Qwen/Qwen3.5-Plus in Qubrid Playground and enter Vision mode.

2. Upload the same origami boat image, enter the exact same English description prompt, and enable Model Reasoning.

3. Save one complete response along with the TTFT, total duration, and token details shown by the page; do not record only the final text.

4. Repeat at least 5 times, then report the mean, standard deviation, and failure count; if switching to the Bailian API, label the provider and model snapshot separately.

Source excerpt or observation (for a compliant short quotation only)

The article's reproducible input is “Describe what you see in this image.”, and it explicitly records a 26.02-second total response, 6.86-second TTFT, and 1,858 reasoning tokens; this article preserves the provider boundary.