We fine-tune open models so they text back like someone you know. And still do the work.
New model
Gemma-4-31B-Humanlike-Chat
Our new open model texts naturally, stays in character, and knows when to use tools.
Our models
Qwen3.8-27B-Humanlike-Chat
Our first open model. It reads the moment: sometimes you need a smart answer, sometimes just someone on your side. It jokes, explains things simply, and handles emails and tool calls when you ask.
| Built on | Qwen3.8-27B |
|---|---|
| Runs on | One 24 GB GPU |
| Formats | GGUF, safetensors |
| Licence | Apache 2.0 |
| Mistaken for the real person | 23.5% (official Qwen 15.1%) |
| Follows instructions it never saw | 43.7 (base 37.3) |
| Asks before it guesses | 58 (base 48) |
| Downloads, last 30 days | 94k |
Blind test: 50% would mean the judge can’t tell. IFBench and When2Call: base is the model it is trained from.