Three model cards, three evaluation questions
An initial evaluation shortlist across text, speech, and multimodal tasks. These are selected models, not a ranking or a list of new September releases.
Start with the task
A model card establishes what the publisher describes. Your evaluation establishes how the candidate handles your work.
Source review: 2026-09-26. This issue contains no Sophrono benchmark results.
| Model | Task | License / terms | Hardware evaluation |
|---|---|---|---|
| Qwen3-8B | Text generation | Apache 2.0 | Measure memory and latency at the selected precision and context length. |
| Qwen3-ASR-1.7B | Speech recognition | Apache 2.0 | Benchmark the supported inference configuration on representative audio. |
| Gemma 3n | Multimodal input; text output | Gemma Terms of Use | Designed for low-resource devices; validate the chosen configuration. |
Read the primary sources
Qwen3-8B
Does the chosen reasoning mode meet the workload quality and latency criteria?
Qwen3-ASR-1.7B
How does transcription perform on your audio, language, and terminology?
Gemma 3n
Which input modalities and quality thresholds does your workflow need?
License review
The linked Qwen cards list Apache 2.0. The Gemma card links its Terms of Use. Record the applicable model version and terms before deployment.
Super Intelligence Newsletter
The frontier of AI, once a month.
The models, research, and releases that change what a business can build, with the sources and the question worth testing.
Test a model on your work