AI that runs on your Mac

Chat, write and work on code with open models
running locally in Particle

Download for Mac View source

Particle's Mac interface, showing a conversation, saved chats and the model selector. — The Mac app, shown with an example conversation

Chat and code, locally

Churo Chat

Handles conversations, writing and questions. Choose Instant or Thinking. Keep chats and optional memories on your Mac.

Particle Code

Works in your project folder. Read and edit files, review changes and run commands with the required approvals.

Choose your model

Install from the app. Adjust experts and memory settings for each model.

Default

Sarvam 30B

The original model behind Churo Chat and Churo Code

Q4_K_M · 19.6 GBView model

Early-access Chat

Nemotron 3.5 Lightning

NVIDIA Nemotron 3.5 Lightning 30B-A3B, available for early-access Chat

MLX 4-bit · ChatView model

Early-access Chat

GPT-OSS-20B

GPT-OSS-20B in mixed MXFP4 / Q8 precision, available for early-access Chat

MXFP4 / Q8 · ChatView model

Early-access Chat

Qwen3.6-35B-A3B

Qwen3.6-35B-A3B, available for early-access Chat

MLX 4-bit · ChatView model

More supported models 10

Qwen3-Coder-30B-A3B MLX 4-bit GLM-4.7-Flash MLX 4-bit Qwen3 Thinking 2507 MLX 4-bit Qwen3 Instruct 2507 MLX 4-bit Gemma 4 26B-A4B MLX 4-bit Qwen3-30B-A3B MLX 4-bit DeepSeek Coder V2 Lite MLX 4-bit Qwen1.5-MoE-A2.7B MLX 4-bit · Fast DeepSeekMoE-16B-Chat MLX 4-bit · Fast MiniCPM-MoE-8x2B Original BF16 · Fast

Alternative runtimes are early-access Chat. Fast is a collection label, not a speed guarantee. MiniCPM uses its larger original BF16 checkpoint.

Sarvam on an M2 with 8 GB RAM

Measured with an internal SSD on a plugged-in Mac

Measured decode speed: Up to 4.3 tokens/s
Peak sampled engine footprint: ~1.8 GiB

Experimental three-expert Sarvam profile, 2K context, short code task. Reduced experts can affect quality; speed varies by workload. Engine footprint excludes system and file-cache memory.

Measurement details

Two naturally completed code responses measured 4.198 and 4.298 decode tokens/s, with 347 post-prefill steps each. The combined rate was 4.247 tokens/s. Decode timing includes reasoning and excludes model loading and prompt prefill.

The recorded peak sampled engine footprint was 1.79 GiB. A separate mixed-language test was slower and reached its reasoning limit without a final answer. These results do not establish the same speed or memory use for every task, longer context or other models. Six experts remains the native model profile.

Recorded measurements

Get Particle

Paste into Terminal to install or update. No Xcode or Python required.

curl -fsSL https://github.com/MathObsession/Particle/releases/latest/download/install.sh | /bin/bash

Apple silicon · macOS 26 or newer

Release notes

Before installing

The app is free to download; models retain their own licenses. Allow about 22 GB free for the default Sarvam installation, plus app space. Memory requirements vary by model and context. Quit Particle before updating; the installer preserves your models, chats and settings. The current tester build is ad-hoc signed and not Apple-notarized.

Local use and privacy

Inference, conversations and optional memories stay on your Mac. Once a model is installed on a connected drive, local chat and file tools work offline. Downloads and updates use GitHub and Hugging Face; approved web tools contact external services. The website uses no analytics or tracking cookies.