How should you compare a dense model with a mixture-of-experts deployment?
This page compares the older editions named in its title. For current GPT-6, Claude 5.5 and Gemini 3.8 comparisons, open the current comparison library. The historical model identities are preserved for migration and reproducibility.
A dense text model with thinking and non-thinking modes.
A dense deployment has different memory and serving requirements from an MoE model. Check the actual quantization and context configuration.
Official source ↗A mixture-of-experts text model in the Qwen3 family.
Active parameters do not equal the total memory needed to load all weights. Compare serving configurations explicitly.
Official source ↗Exact checkpoint fields from pinned publisher files. The configured position limit does not establish the context enabled by a serving endpoint.
| Configuration field | Qwen3 32B | Qwen3 235B A22B |
|---|---|---|
| Published architecture | Qwen3ForCausalLM | Qwen3MoeForCausalLM |
| Transformer layers | 64 | 94 |
| Configured position limit | 40,960 | 40,960 |
| Pinned configuration revision | 9216db5781bf | 8efa61729e24 |
| Configuration checked | 2026-10-10 | 2026-10-10 |
| Original configuration | Read pinned file ↗ | Read pinned file ↗ |
An MoE active-parameter label does not describe all resident weights. Compare complete hosted request costs or complete deployment costs, then test the tasks that justify the larger system.
Use the same representative inputs and a written answer rubric. Pin the model version and record the endpoint, thinking configuration, quantization when known, context length, output allowance and concurrency. Preserve native request differences in an adapter rather than silently dropping fields.
Measure full request time, error rate and reported usage. For tools, validate arguments and score the result after execution. For images, preserve resolution and preprocessing. A valid response format alone does not establish a correct answer.
Calculate input cost plus all billed output, including reasoning where the provider counts it. Add retries and discarded answers. Divide the total by the number of results that pass your quality criteria. Keep API usage cost separate from deployment and engineering cost.
Estimate token costs with the API calculator →Enter results from your own evaluation. A task should count once it meets your written quality criteria. Include failed attempts and retries in total spend. These inputs stay in this browser and are not a Flonno benchmark.
Enter all four values. Attempted tasks must be at least one; accepted tasks cannot exceed attempted tasks.
Enter all four values. Attempted tasks must be at least one; accepted tasks cannot exceed attempted tasks.
A source benchmark can inform your shortlist, while your evaluation shows whether a model finishes the work at an acceptable cost. Keep the chosen endpoint, processing terms and output limits in the decision record. Inclusion in the Radar does not mean the model is served by Flonno.
For the API currently offered here, explore Qwen3.8 Flash 27B and the production evaluation guide.
When is a larger dense Qwen model worth the extra serving cost?
Which reasoning configuration fits your latency and output budget?
Do you need text reasoning, image inputs, or both?
When should you choose open-weight serving over a hosted API?
