Identical weights, identical Mac, identical prompts, shipping defaults on both sides — mlx-serve decodes +26% faster (geomean) with +36% prefill on MLX models. Native menu-bar app, MIT license, no Electron, no Python.
Shipping defaults on both sides, same machine, same prompts, identical MLX weight files. Reproduce with tests/bench.sh --family all --lmstudio --omlx --mtplx — the harness ships in the repo.
| Capability | mlx-serve | LM Studio |
|---|---|---|
| MLX + GGUF models | ✓ | ✓ |
| Decode speed, identical weights (defaults) | +26% geomean | baseline |
| OpenAI-compatible API | ✓ | ✓ |
| Anthropic Messages API (Claude Code) | ✓ | partial |
| OpenAI Responses API + WebSockets | ✓ | partial |
| Ollama API (drop-in for Ollama clients) | ✓ | ✗ |
| Speculative decoding (PLD + drafter + MTP) | ✓ | ✗ |
| KV-cache quantization | ✓ | ✗ |
| Continuous batching | ✓ | ✗ |
| Agent mode + MCP client | ✓ 10 tools | ✗ |
| Sandboxed agent shell (Linux VM) | ✓ | ✗ |
| DeepSeek V4 Flash (284B) | ✓ via ds4 | ✗ |
| Image / video / voice generation, local | ✓ | ✗ |
| App runtime | Native Swift | Electron |
| License | MIT | Proprietary |
Feature set as of v26.8.7. Benchmark details, CSVs, and the reproduction harness live in the repo.
MLX Core reads LM Studio's own ~/.lmstudio/settings.json and auto-discovers your existing model folder, so everything you've already downloaded shows up in the picker immediately — MLX and GGUF alike. Your OpenAI-compatible clients keep working too: same wire protocol, just a different port.
Yes. On the v26.8.3 matrix (M4 Max, LM Studio 0.4.19+2, identical MLX weight files, both engines on shipping defaults) mlx-serve decodes +26% geomean and prefills +36% geomean across the four models LM Studio also has. The separation is prefill (+117% on E4B) and speculative decoding: on Qwen 3.6 27B mlx-serve loads the checkpoint's MTP head and LM Studio does not, which is +145% on the same file. The benchmark harness ships in the repo so you can reproduce it on your own machine.
No. MLX Core auto-discovers LM Studio's model folder via ~/.lmstudio/settings.json — everything on disk appears in the picker. A custom-folder picker covers models stored anywhere else.
Deeper Anthropic Messages (Claude Code works natively) and OpenAI Responses coverage than LM Studio’s newer compatibility endpoints — including a WebSocket transport and response compaction — a drop-in Ollama API, agent mode with MCP, an isolated Linux VM for agent shell commands, speculative decoding, KV-cache quantization, continuous batching, DeepSeek V4 Flash, and fully local image, video, and voice generation.
Yes — MIT license, server and app both. LM Studio is proprietary freeware; mlx-serve you can read, fork, and ship.
Download MLX Core — it finds your LM Studio models on first launch, so the switch costs you nothing but the download.