@ivanfioravanti Very confusing name.
It’s meant to assist Gemma (by narrowing the scope of possible next tokens during generation), not the human. It’s also 400 million parameters, not 12 billion.
@Jaaneek So is xAI fine with using the Grok Build OAuth client from third-party harnesses in general (in an honest way, so with a custom User-Agent set)?
Or is your tolerance limited to those three clients (OpenCode, OpenClaw, Hermes) specifically?
I was digging into this and I’m amazed by the way it works.
OpenCode doesn’t have a valid OAuth app, so they use the sign-in flow for Grok Build. Same as Hermes and OpenClaw.
Update on this:
I went looking for some simpler models that are working out of the box and still match my criterias (20–40 billion parameters, dense, agentic, preferably not multimodal, preferably not open-source) and found this one.
Seems perfect for my current purposes.
Burnt $ 100 on tokens today, got barely any value out of it, decided to take a small break from cloud inference, asked GPT-5.5 to prepare 8 different inference server setups for my local rig, all of them failed.