README.md

metadata

title: Pocket Pikachu — A Pokémon You Talk To, On-Device
emoji: ⚡
colorFrom: yellow
colorTo: red
sdk: gradio
sdk_version: 6.17.3
app_file: app.py
pinned: true
license: apache-2.0
short_description: On-device talking Pikachu — MiniCPM5 + VoxCPM voice on GPU
models:
  - openbmb/MiniCPM5-1B-GGUF
  - openbmb/VoxCPM-0.5B
  - Systran/faster-whisper-small.en
tags:
  - track:wood
  - sponsor:openbmb
  - achievement:offbrand
  - achievement:offgrid
  - gradio
  - hackathon
  - build-small-hackathon
  - minicpm
  - voxcpm
  - on-device
  - local-llm

⚡ Pocket Pikachu — a Pokémon you actually talk to (VoxCPM on ZeroGPU)

A clickable, talking digital pet that runs in-process on a single GPU. Click the mic (or type), and your pocket Pikachu hears you, thinks, and talks back — no cloud APIs. Built for the Build Small hackathon: every model is ≤ 4B params.

Stack — in-process on ZeroGPU:

Role	Model	Where it runs
🧠 Brain	openbmb/MiniCPM5-1B-GGUF (Q4_K_M)	OpenBMB, via llama-cpp-python — CPU
🗣️ Voice	openbmb/VoxCPM-0.5B	OpenBMB, cute female ref + pitch↑/tempo↓ styling — GPU
👂 Ears	Systran/faster-whisper-small.en	CTranslate2 int8 — CPU

Why no NeMo/Nemotron here: nemo_toolkit[asr] cannot build on a ZeroGPU Gradio Space — its multi-GB dependency tree plus a CUDA torch OOM-kill the CPU build container (exit 137), and the streaming model has no transformers path. Nemotron runs in the native macOS PocketDM app instead. This Space's upgrade over its free-CPU sibling is VoxCPM (OpenBMB's neural voice) on GPU.

How it works

app.py is a single Gradio Blocks app. Per-turn inference runs inside one @spaces.GPU-decorated function (run_turn_gpu) so the GPU is attached for the turn; models lazy-load on first use. A turn flows:

1. Mic / text in — gr.Audio(sources=["microphone"]) or a textbox.
2. Transcribe — faster-whisper small.en turns speech into text (CPU).
3. Reply — MiniCPM5-1B generates Pikachu's response with the same system prompt + keyless tool-fact grounding (time / weather / web lookup) as the desktop companion (CPU via llama.cpp).
4. Speak — VoxCPM-0.5B synthesizes the reply on the GPU in a consistent cute female voice (bundled reference clip + pitch-up/tempo-down ffmpeg styling), played via gr.Audio(autoplay=True).

Custom (non-stock-Gradio) UI: sunny gradient, bobbing sprite, chat bubbles, daily check-in, hero + "what can it do" sections.

Why it qualifies

• Best MiniCPM Build (sponsor:openbmb) — built on OpenBMB's MiniCPM5-1B (brain) and VoxCPM-0.5B (voice).
• Off the Grid (achievement:offgrid) — no cloud APIs; every model runs locally in the Space process.
• Off-Brand (achievement:offbrand) — fully custom UI beyond stock Gradio.
• Thousand Token Wood (track:wood) — a whimsical talking desktop pet.
• Tiny / ≤4B — MiniCPM5-1B, VoxCPM-0.5B, faster-whisper-small all well under 4B. (The native app adds NVIDIA Nemotron-0.6B for streaming ASR.)

Hardware

Runs on ZeroGPU (zero-a10g) — the free, quota-based GPU available to members of the build-small-hackathon Team-plan org (org-billed, no personal cost). The voice (VoxCPM) runs on the GPU; the brain (llama.cpp) and Nemotron ears on mac or (faster-whisper) run on CPU.