koe-270m
koe (ε£°, "voice") β xvoidsx's family of small voice-assistant models. The voice of the wired, built for Hey Lain and local-first AI on open hardware.
At a glance
| Params | ~270M |
| Quantized | ~200 MB (Q4_K_M) |
| License | Apache 2.0 (intended) |
| Runs on | Raspberry Pi, 2015-era laptops, anything |
Which variant should I pick?
| Variant | Size | For |
|---|---|---|
koe-270m (this) |
~200 MB Q4 | Weakest hardware β the perma-computing floor |
koe-600m (planned) |
~400 MB Q4 | Everyday laptops, navi mini |
koe-1b (planned) |
β | Machines with headroom |
Same voice, different weight classes. Every variant ships as GGUF for ollama:
ollama pull hf.co/xxvoidsxx/koe-270m
How this model was made
In progress β training data (wired-voice conversations) is being curated. This section will document the full recipe: base model, dataset, hyperparameters, and validation, the way research should be documented.
Benchmarks
To come. Validated on real old hardware, because that's the point β if it runs on a 2015 laptop, it runs anywhere.
Built by xvoidsx β researching local AI on open hardware.