Hi Anna team,
I tested the multi-model capabilities provided by Anna in a PPT generation workflow, and noticed a very large latency gap between different models. The Gemini models respond quickly, while several other models are significantly slower, which affects the actual developer experience.
The data below comes from my local workspace logs. I only included successful LLM calls where the log explicitly contains a model field. The main metric is duration_ms, with response._meta.latencyMs used as a reference. Calls without an explicit model field, such as some agent or research tasks, are excluded.
| Model | Successful Calls | Avg Duration | Median | Min | Max | Avg Provider Latency |
|---|---|---|---|---|---|---|
google/gemini-3.1-flash-lite-preview |
16 | 4.8s | 4.4s | 2.4s | 8.6s | 4.1s |
google/gemini-3-flash-preview |
4 | 7.1s | 6.0s | 4.2s | 9.4s | 6.2s |
xiaomi/mimo-v2.5-pro |
1 | 42.6s | 42.6s | 42.6s | 42.6s | 40.8s |
qwen3.7-plus |
3 | 44.5s | 41.8s | 16.3s | 75.5s | 43.4s |
minimax/minimax-m3 |
1 | 73.3s | 73.3s | 73.3s | 73.3s | 71.3s |
moonshotai/kimi-k2.6 |
4 | 75.2s | 63.7s | 62.9s | 98.9s | 74.4s |
Based on this data, the Gemini models generally respond within 5-7 seconds, while several other models often take 40-75 seconds, with some calls approaching 100 seconds. This is not a minor fluctuation; it is a major difference in interactive usability.
For smaller calls such as suggest_context, the gap is also clear:
| Model | suggest_context Duration |
|---|---|
google/gemini-3.1-flash-lite-preview |
~3.0s avg |
google/gemini-3-flash-preview |
4.2s |
qwen3.7-plus |
16.3s |
xiaomi/mimo-v2.5-pro |
42.6s |
minimax/minimax-m3 |
73.3s |
moonshotai/kimi-k2.6 |
75.4s |
My main concern is not model quality, but the very large latency difference between the model options currently exposed through Anna. For interactive generation workflows, such as outline generation, page planning, and research planning, waiting tens of seconds or more than a minute breaks the flow and can make the task feel stuck or failed.
Could the Anna team help confirm:
- Is this level of latency expected for the non-Gemini models?
- Could there be extra overhead from routing, provider queues, streaming, timeout behavior, or Anna’s proxy layer?
- Could the model selection UI show an approximate speed tier, so users can avoid choosing very high-latency models unintentionally?
- Could slower models have clearer progress indicators, timeout messages, or fallback behavior?
- Could the response path for models such as
qwen3.7-plus,moonshotai/kimi-k2.6, andminimax/minimax-m3be prioritized for optimization?
From the logs, Gemini is clearly usable for interactive workflows. However, the latency of several other models is too high, making the multi-model experience feel unbalanced. I hope the team can investigate and optimize this.