Rapid-MLX 1.0: a built-in AI agent and new local models for Mac

Run AI on Apple Silicon with a Desktop app, a terminal agent, and OpenAI- and Anthropic-compatible APIs. Start with one local chat; connect more tools when you need them.

Rapid-MLX runs AI models locally on your Apple Silicon Mac. Version 1.0 adds a built-in agent harness, LFM3 text and vision models, Ornith 122B, and MiniMax-H3 video with audio. The Desktop app and engine are free and open source under Apache 2.0. No account is needed for local inference.

Here is the shortest route from download to a useful conversation.

1. Install

Download Desktop for Apple Silicon, macOS 14 or later. Open the app and choose a model that leaves memory for macOS and your other apps.

Prefer Terminal? The installer sets up the engine:

curl -fsSL https://rapidmlx.com/install.sh | bash

Restart Terminal if the command is not on your PATH. The built-in harness needs Python 3.11 or newer. See installation help.

2. Chat with a local model

In Desktop, select a model and start a chat. In Terminal, try the small LFM3 text model:

rapid-mlx chat lfm3-1.7b-8bit

First use downloads weights; allow time and disk space. LFM3's download is about 1.84 GB, but runtime memory includes more than weights. Ask it to explain a short piece of code or draft a message. Model quality and speed depend on the model, prompt and Mac; pick for your memory.

3. Give it a task

To connect an installed coding agent, use the agent guides. For example:

rapid-mlx launch codex

Or start the built-in agent from the folder you want it to work in:

rapid-mlx chat --harness

It can read files, search your workspace, edit files, run commands and Python, and use web tools. It shows tool and model-loading progress, supports image attachments with vision models, and keeps resumable local history. Use --approvals always to review every action or --no-action-tools to disable edits and execution. Read the harness guide before handing it a workspace.

Terminal commands are not sandboxed. The restricted environment and approval patterns reduce accidental exposure; commands still have your local account's access. Network-client commands trigger approval heuristics; built-in web tools send requests externally and do not ask per domain. Keep sensitive work out of reach and review actions.

What's new in 1.0

• LFM3 1.7B and LFM3-VL 2B: pinned 8-bit text and image aliases, also offered in Desktop. Models and memory details.
• Ornith 1.5 122B-A10B: a pinned 4-bit text alias. The engine sets a 96 GiB unified-memory floor; weights alone are about 69.61 GB. Longer contexts and concurrency need more headroom. Ornith family.
• MiniMax-H3: text or a first image to video with synchronized audio. rapid-mlx serve minimax-h3 needs 160 GB+ unified memory, BF16 weights, and defaults to the 8-step Turbo LoRA. Video requirements.

Desktop

Attach up to eight images per message, optionally release an idle model after your chosen timeout, and use a more compact agent connection page. Share Compute has a simpler key → model → start-sharing flow; it remains experimental.

Desktop's Built-in Agent Harness is experimental and off by default. Enable it in Settings for new chats. It supports images; the familiar chat flow remains available. Harness behavior and limits.

Measurements

The release measured LFM3 on an M5 Max, 36 GB, macOS 27.0.1, MLX 0.32.4, Python 3.12.13, source 46ac39d53: median of five after warmup, greedy decoding, thinking off, 256 output tokens for text.

Vision answers were only 35 tokens, so that decode figure is a short-answer estimate. Peak reported Metal memory was 6.81 GB for text and 3.24 GB for vision, not whole-system memory requirements. The machine had background activity; these results are workload-specific.

Eight concurrent text requests reached 951.5 aggregate tok/s, each generating 256 tokens, measured as total completion tokens / batch wall time including prefill and client overhead. This is aggregate throughput, not one chat's speed. See the release measurements and per-chip speed guide.

Upgrade and limits

rapid-mlx upgrade

Homebrew users can run brew upgrade rapid-mlx; pip users can run pip install -U rapid-mlx==1.0.0.

Long tasks still need context management: narrow file reads, keep output short, and use /context and /compact in the harness. Small models can stop using tools over long conversations. Compaction cannot expand a model's context window. The long-task guide explains recovery.

Experimental Qwen3.8-27B TensorFold sharing supports text and tools, but rejects response_format, grammar and media inputs. See all release caveats.

Frequently asked questions

Does my chat leave my Mac?

Inference runs locally. Online tools send requests to external services. For offline harness use, start a new owner with RAPID_MLX_CHAT_WEB_TOOLS=0 rapid-mlx chat --harness --profile offline.

Do I need a large Mac?

No. Start with a small model and leave system memory headroom. Ornith's 96 GiB memory floor and MiniMax-H3's 160 GB+ unified-memory requirement are separate, large-machine options.

Are agent commands sandboxed?

No. Terminal and Python run with your local account's access. Review approvals, use --approvals always, or disable action tools with --no-action-tools. Desktop harness integration is experimental and off by default.

curl -fsSL https://rapidmlx.com/install.sh | bash