Qwen 3

Qwen/Qwen3-0.6B-MLX-4bit

Vendor: Alibaba
Quantization: MLX 4-bit
Parameters: 600M
Size: 313.3 MB

$ brew install mirai$ mirai --model Qwen/Qwen3-0.6B-MLX-4bit

Benchmarks

0.13 s

lower is better ↓

0.13 s

uzu

0.23 s

mlx

0.20 s

llamacpp

473 t/s

higher is better ↑

473 t/s

uzu

463 t/s

mlx

264 t/s

llamacpp

0.82 GB

lower is better ↓

0.82 GB

1.95 GB

0.70 GB

0.31.2 llama.cpp0.3.23/1,312 input tokens/512 output tokens

Benchmarked 7 Aug 2026

Integrate with SDK

Choose framework

Run the following command to install Mirai SDK

https://github.com/trymirai/uzu-swift

Apply code

1import Foundation2import Uzu34public func runChat() async throws {5    let engineConfig = EngineConfig.create()6    let engine = try await Engine.create(config: engineConfig)7    8    guard let model = try await engine.model(identifier: "alibaba:qwen3.5:0.8b:mirai:mirai-m:4") else {9        return10    }11    for try await update in try await engine.download(model: model).iterator() {12        print(String(format: "\r\u{001B}[2KDownload progress: %.2f%%", update.progress() * 100), terminator: "")13        fflush(stdout)14    }15    print()16    17    let messages = [18        ChatMessage.system().withText(text: "You are a helpful assistant"),19        ChatMessage.user().withText(text: "Tell me a short, funny story about a robot")20    ]21    let session = try await engine.chat(model: model, config: .create())22    let stream = await session.replyWithStream(input: messages, config: .create())23    var message: ChatMessage? = nil24    for try await update in stream.iterator() {25        switch update {26        case .replies(let replies):27            let reply = replies.last28            message = reply?.message29            print("Generated tokens: \(reply?.stats.tokensCountOutput ?? 0)")30        case .error(let error):31            print("Error: \(error)")32        }33    }34    print("Reasoning: \(message?.reasoning() ?? "empty")")35    print("Text: \(message?.text() ?? "empty")")36}37

1Choose framework

2Run the following command to install Mirai SDK

https://github.com/trymirai/uzu-swift

3Apply code

1	import Foundation
2	import Uzu
3
4	public func runChat() async throws {
5	let engineConfig = EngineConfig.create()
6	let engine = try await Engine.create(config: engineConfig)
7
8	guard let model = try await engine.model(identifier: "alibaba:qwen3.5:0.8b:mirai:mirai-m:4") else {
9	return
10	}
11	for try await update in try await engine.download(model: model).iterator() {
12	print(String(format: "\r\u{001B}[2KDownload progress: %.2f%%", update.progress() * 100), terminator: "")
13	fflush(stdout)
14	}
15	print()
16
17	let messages = [
18	ChatMessage.system().withText(text: "You are a helpful assistant"),
19	ChatMessage.user().withText(text: "Tell me a short, funny story about a robot")
20	]
21	let session = try await engine.chat(model: model, config: .create())
22	let stream = await session.replyWithStream(input: messages, config: .create())
23	var message: ChatMessage? = nil
24	for try await update in stream.iterator() {
25	switch update {
26	case .replies(let replies):
27	let reply = replies.last
28	message = reply?.message
29	print("Generated tokens: \(reply?.stats.tokensCountOutput ?? 0)")
30	case .error(let error):
31	print("Error: \(error)")
32	}
33	}
34	print("Reasoning: \(message?.reasoning() ?? "empty")")
35	print("Text: \(message?.text() ?? "empty")")
36	}
37

Details

A compact, 4-bit quantized version of Qwen3-0.6B optimized for Apple Silicon via the MLX framework. Published by the Qwen team, this model brings the latest-generation Qwen3 capabilities to efficient on-device inference on Mac hardware.

Key Features

Qwen3-0.6B-MLX-4bit inherits the standout capabilities of the Qwen3 family in a remarkably small footprint:

Dual-mode reasoning: Seamlessly switch between a thinking mode (for step-by-step logical reasoning, math, and code) and a non-thinking mode (for fast, general-purpose dialogue) — all within a single model. Users can toggle behavior via `enable_thinking` or inline `/think` and `/no_think` tags in conversation.
Multilingual support: Covers 100+ languages and dialects with strong instruction-following and translation performance.
Agent & tool-calling capabilities: Designed for integration with external tools, compatible with frameworks like Qwen-Agent, and capable of handling complex agentic workflows.

Architecture

Type: Causal language model (dense transformer)
Parameters: 0.6B total (0.44B non-embedding)
Layers: 28, with grouped-query attention (16 Q heads, 8 KV heads)
Context length: 32,768 tokens
Quantization: 4-bit (MLX format)
Base model: Qwen3-0.6B-Base, with full pretraining and post-training

Intended Use

This model is ideal for developers building lightweight, privacy-friendly applications on Apple Silicon — think local chatbots, coding assistants, creative writing tools, or multilingual agents — where low latency and minimal memory consumption matter. It requires `mlx_lm` ≥ 0.25.2 and `transformers` ≥ 4.52.4.

Provenance

Developed and released by the Qwen team under the Apache 2.0 license. Full technical details are available in the Qwen3 Technical Report.

Explore all local models