Kognitiv Playground

Kognitiv Playground is a free playground for running large language models entirely on your Android phone — no account, no sign-in, and no cloud. Every token is generated locally on your device's CPU. It's built for ML engineers, researchers, CS students, and the local-LLM community who want to poke at how modern on-device AI actually works.

This is a technical research demo, not a chatbot assistant. Two hands-on experiments ship today:

▸ BLOCK DIFFUSION LLM
Most language models write left to right, one token at a time. This one doesn't. It runs BD3LM-Qwen3-0.6B — a block-diffusion language model — on device through Google's LiteRT-LM runtime. Instead of predicting the next word, it denoises masked blocks of tokens by confidence, so you can literally watch the answer resolve out of noise. Ask "What is the capital of France?" and see "Paris" settle in a few seconds on a flagship device. The model (600 MB) downloads on demand from Google Play the first time you open the experiment.

▸ TEMPORALKV PLAYGROUND
As a conversation grows, an LLM can't keep every past token in memory — it has to decide what to remember and what to evict from its KV cache. This experiment runs Qwen2.5-1.5B-Instruct (Q4_K_M) on device via llama.cpp and visualizes, live, which past tokens are kept versus evicted under different cache-eviction policies — including a learned policy called TemporalKV. It's an interactive window into active research on making long-context on-device LLMs cheaper. The model (1 GB) downloads on demand.

WHY ON-DEVICE
Because inference runs locally, your prompts and conversations never leave the phone. There's no server, no telemetry on your text, no login. Offline AI that's genuinely private by construction — useful if you care about data privacy, latency, or just understanding how these systems run without a datacenter.

FEATURES
• 100% on-device inference — CPU, no network calls for generation
• Two experiments: block diffusion LLM + KV-cache eviction visualizer
• Watch a diffusion LLM denoise an answer in real time
• See live which tokens an LLM keeps vs evicts from its KV cache
• Powered by LiteRT-LM and llama.cpp
• Models download on demand from Google Play (~600 MB and ~1 GB)
• Modern Jetpack Compose + Material 3 interface
• No account, no ads, no in-app purchases
• Android 8.0+ on arm64-v8a devices

A NOTE ON SCOPE
These are small models (0.6B and 1.5B parameters). They're great for demonstrating on-device techniques but are not as capable as large cloud assistants, and they can be wrong — don't rely on them for factual, medical, financial, or legal advice. TemporalKV is an evolving research demo, not a finished product. The point is to make on-device LLM internals visible and tinkerable.

Free. Private. Local. Happy tinkering.