| ``` +--------------------------------------------------+ | yuy-chat v0.1.0 Yuuki-best | Balanced | +--------------------------------------------------+ | | | You: Explain async/await in Rust | | | | Yuuki: async/await in Rust allows you to write | | asynchronous code that looks synchronous. The | | async keyword marks a function as returning a | | Future, and await suspends execution until the | | Future resolves... | | | +--------------------------------------------------+ | Message: _ | +--------------------------------------------------+ | Enter: Send | Ctrl+C: Menu | Ctrl+S: Save | +--------------------------------------------------+ ``` |
**A full chat experience in your terminal.**
Select models interactively. Stream responses word by word. Save and reload conversations. Switch presets on the fly. All running locally on your machine. All powered by ratatui + Rust. |
Interactive ChatReal-time streaming responses displayed word by word. Multi-line input with Shift+Enter. Scrollable message history with keyboard navigation.Model SelectorAuto-discovers `.gguf` and `.llamafile` models from your local directory. Navigate with arrow keys, select with Enter. Refresh without restarting.Conversation HistorySave conversations as JSON files. Load previous chats from a built-in conversation browser. Delete old sessions you no longer need.HuggingFace CloudOptional API integration for cloud-based inference. Configure your HF token in the settings screen. Local and cloud models appear side by side in the selector. |
Generation PresetsThree built-in modes -- Creative (0.8 temp), Balanced (0.6 temp), and Precise (0.3 temp). Cycle between them with a single keypress. Custom presets planned for v0.2.Settings ScreenConfigure models directory, HuggingFace token, default preset, history saving, and UI theme -- all from within the TUI.Cross-PlatformRuns on Termux (Android), Linux, macOS, and Windows. Same binary, same interface, same experience. Mobile-first defaults for constrained hardware.Lightweight~8 MB binary. ~20 MB idle RAM. ~50 ms startup. Built with Rust for zero-overhead performance and memory safety. |
Model Formats| Format | Extension | Notes | |:-------|:----------|:------| | GGUF | `.gguf` | Recommended. Requires llama.cpp | | Llamafile | `.llamafile` | Self-executing. Zero dependencies | |
Inference Runtimes| Runtime | Type | Notes | |:--------|:-----|:------| | llama.cpp | Local subprocess | Default. Fast, CPU-optimized | | llamafile | Local executable | Bundled runtime + model | | HuggingFace API | Cloud HTTP | Optional. Requires token | |
| **Training Details** | | | |:--|:--| | Base model | GPT-2 (124M parameters) | | Training type | Continued pre-training | | Hardware | Snapdragon 685, CPU only | | Training time | 50+ hours | | Progress | 2,000 / 37,500 steps (5.3%) | | Cost | $0.00 | | **Quality Scores (Checkpoint 2000)** | Language | Score | |:---------|:------| | Agda | 55 / 100 | | C | 20 / 100 | | Assembly | 15 / 100 | | Python | 8 / 100 | |