How it compares

mlx-bun uses MLX through mlx-c. It brings model execution and serving into Bun so TypeScript applications can integrate inference directly or distribute a local executable.

mlx-lm and mlx-optiq supply reference implementations for the project’s numerical comparisons. Their documentation describes their current capabilities. mlx-bun’s supported subset and extensions are documented in its model roster.

Choosing an integration

Use the library API for direct Bun integration, an isolated host for a desktop application’s local worker, or the HTTP API for clients in another runtime. Protocol compatibility covers the documented routes and fields; it does not promise every feature of a hosted provider.

Continuous batching, speculative methods, quantized cache layouts, and RAM/SSD reuse compose where the model implementation supports them. The configuration reference records those combinations and the explicit serial option. Training and model conversion have separate training and CLI documentation.

Compare evidence for your workload

The benchmark ledger records complete requests, startup, prefill, decode, cache reuse, and concurrent throughput. Read each result with its machine, artifact, settings, and limits. A win on one workload does not imply a win on every model or context length.

The correctness contract distinguishes stock mlx-lm parity, mlx-optiq extensions, and experimental methods. Quantizing a cache changes the computation; the appropriate oracle is the matching scheme.