中文 · English
Compare AI models by answer quality, speed and cost in one local-first workspace. Runs on Windows, macOS, Linux and Android, against cloud APIs or local model servers.
- Model arena — send one prompt to several models and compare streaming answers side by side. Each model keeps its own conversation history; image input and generation work where the model and API support them.
- Batch experiments — combine test sets, models, parameter groups and repetitions. Every case runs independently with a saved configuration snapshot, and you can pause, resume or retry failures.
- Answer evaluation — rate answers manually, check expected answers with exact, substring or JSON matching, or let a judge model score anonymous answers for accuracy, instruction following and completeness.
- Structured decisions — define questions and criteria for judgment, choice and scoring tasks, then compare decisions, probabilities and confidence across models.
- Performance and cost — time to first token, generation speed, duration, success rate, token usage, cache hit rate and cost. API-reported values are kept distinct from estimates.
- Models and materials — OpenAI, Anthropic, Gemini, DeepSeek, OpenRouter and other providers, custom OpenAI-compatible endpoints, and local Ollama or LM Studio servers. Prompts, test sets and parameter presets are reusable.
- Reports and backups — export Markdown, HTML, CSV or JSON, and back up or restore the workspace. Records stay on your device; model requests go only to the services you configure.
- Download the package for your platform from Releases.
- On Models, add a provider with its endpoint, key and models. For local models, start your Ollama or LM Studio server first.
- Select models in the Arena and send a prompt. For batch evaluation, prepare cases under Tests, then create an Experiment.
Cloud models use your own credentials and are billed by the provider. Image support, sampling parameters and usage metrics depend on the model and the API.
Tauri 2, Rust, React 19, TypeScript, Zustand and Vite. Requires Node.js 24+, pnpm 11 and Rust stable; Linux also needs WebKitGTK 4.1 and the usual Tauri system packages.
pnpm install --frozen-lockfile
pnpm tauri:dev # desktop app
pnpm dev # frontend in a browser
pnpm check # types, lint and unit tests
pnpm test:e2e # end-to-end tests (Playwright)
pnpm tauri build # desktop packages| Path | Contents |
|---|---|
src/lib |
Bottom layer: types, validation, provider catalog and transport, generation config, statistics, store and storage |
src/features/<x> |
One directory per feature: domain/ is pure logic, components/ and hooks/ are UI |
src/pages |
Route entry points, loaded lazily |
src/app |
Router, navigation, providers, global hotkeys, queue processor |
src-tauri |
Rust shell: plugin registration and Linux AppImage environment setup |
Two Web Workers keep long-running work off the UI thread: generation streams (src/lib/workers) and Markdown parsing (src/features/chat-arena/markdown). The generation worker is an entry of the main bundle, so the window and the worker share a single copy of the provider SDK.