Team Ai
Apppublic

BananaMind/BananaMind-CodeQ-WebGPU-Pi

sourceHugging Faceapache-2.0updated 18d agoView on Hugging Face
0likes
App README

Pi browser workspace · BananaMind-CodeQ-1.3-2B

Run Pi's agent core, just-bash, and [BananaMind-CodeQ-1.3-2B](https://huggingface.co/BananaMind/BananaMind-CodeQ-1.3-2B) entirely in your browser. Chat with Pi, inspect and edit files, and run shell commands in a persistent virtual workspace. Inference happens locally on WebGPU — no API key, no server.

This is a remix of victor/MiniCPM5-2B-WebGPU-Pi that swaps in BananaMind-CodeQ-1.3-2B, a LoRA fine-tune of MiniCPM5-2B on agentic coding traces (Pi, GLM, DeepSeek and other agent/debugging traces).

Model

Weights come from BananaMind/BananaMind-CodeQ-1.3-2B-ONNX (pinned revision 0ef7af717534c4a3c12675e7d65c930e6ce1bb96). Because CodeQ has the same architecture as MiniCPM5-2B, its weights were written directly into the optimized q4f16 graph from the original Space (fused QKV, GQA, last-position LM head, split embeddings): symmetric int4, block 32, fp16 activations and KV cache. About 1.84 GB on first download, SHA-256 verified and cached in OPFS.

Requirements & limits

WebGPU with shader-f16, OPFS, and enough memory. 8,192-token context budget, up to 2,048 generated tokens per call, 12 turns per request. If storage is blocked in the embedded view, use Open workspace in a new tab.

Pi has bash, read, write and edit tools; just-bash supports pipes, redirects, awk, jq, sed, grep, etc. No Node.js, Python, network or native binaries.

Attribution

Base model OpenBMB MiniCPM5-2B (Apache-2.0), fine-tune BananaMind, ONNX graph Mike0021/MiniCPM5-2B-ONNX, agent Pi / earendil-works, shell just-bash / Vercel, inference Transformers.js and ONNX Runtime Web. Original app by the authors of victor/MiniCPM5-2B-WebGPU-Pi. License texts are in licenses.