A lot of people want to use small models (< 10 GB) to drive assistants. The needs here are different from those associated with large models.
For assistant use, we want models that are small, quick, and reliable rather than those that can solve world-changing problems and write entire codebases.
To try to understand which models are currently (Nov 2026) best, I created EverydayBench, a benchmark suite tuned towards everyday tasks such as summarization, translation, and image analysis (e.g., reading a bus timetable). This benchmark suite intentionally excludes coding and long-horizon tasks. I then ran it on LFM 3 Omni, Gemma E4B, and Nemotron 3.5 E4B.
At only 2.3 GB, LFM 3 Omni Q8 performed as well as the larger Gemma E4B Q8 on summarization and was good enough on image analysis. It did well on analysis of short audio snippets (1 and 5 min), but failed on 30 min snippets. LFM also couldn’t match Gemma on translation.
In the end, that’s not bad for a model that takes less than a third of the space of Nemotron 3.5 Q4KM and Gemma E4B. However, you have to know that LFM’s translation capabilities don’t match up.
I’ve tried a bunch of 2B-12B models to use in home assistant to support the voice pipeline and found Gemma E4B worked best. No real benchmarks, just based on vibes from manual testing. Felt it gave a good balance on executing commands correctly and response time.