VIBETHINKER · WEBGPU TRAIN + HOT-SWAP LoRA

Train VibeThinker-3B in your browser — full WebGPU backward pass + AdamW over the int4 base — then hot-swap the LoRA into inference live. No server. No upload.
▶ Inference
✦ Train (LoRA)
⚙ Inference is paused while VibeThinker-3B is learning in the Train tab…
Model not loaded Press “Load VibeThinker-3B” to begin. First load streams the weights once, then it's cached.
Step 1 · Install the model
~6 GB, one-time — streamed from Hugging Face & cached in your browser. Other sources under ⚙.

Forge a skill for your knife — a LoRA trained entirely in this tab that compiles a plain request into a precise macro over a small, typed action space. This is constrained codegen (the model's strength), not chat. Every skill saves to your knife and hot-swaps into Inference.

⚔ Inbox & Calendar skill

Teaches the model to turn requests like “email my mom and book a reminder to respond” into a verifiable macro over a fixed set of blades — compose_email, reply_email, create_event, set_reminder, find_slot… — and to bounce anything outside inbox/calendar with OUT_OF_SCOPE. The action space is supplied, so it never invents an API; the emitted macro is checkable.

    Forge a custom skill from your own text

    Paste the private text you would never send to a cloud model — notes, decision rules, your own action vocabulary — and the skill is trained locally. For strongest results, include explicit rules, examples, or acceptance criteria the model can reason over. Everything stays local to this browser session unless you export it.

    ① Build & tokenize
    ② Forward
    ③ Backward (grads)
    ④ AdamW update
    ⑤ Hot-swap live
    Done. The tuned adapter is live and saved to your fine-tunes. Open it in Inference.
    ⚔ KNIFE
    In-browser WebGPU training · full backward + AdamW · runtime LoRA hot-swap. 🜂 · model · source