Gemma-3

Name: mlx-community/gemma-3-27b-it-4bit
Author: Google

mlx-community/gemma-3-27b-it-4bit

Vendor: Google
Quantization: MLX 4-bit
Parameters: 27B
Size: 14.2 GB

$ brew install mirai$ mirai --model mlx-community/gemma-3-27b-it-4bit

Benchmarks

5.33 s

lower is better ↓

5.33 s

uzu

5.71 s

mlx

5.59 s

llamacpp

26 t/s

higher is better ↑

26 t/s

uzu

28 t/s

mlx

24 t/s

llamacpp

16.38 GB

lower is better ↓

16.38 GB

18.99 GB

15.88 GB

0.31.2 llama.cpp0.3.23/1,343 input tokens/299 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 4-bit quantized, MLX-optimized conversion of Google's Gemma 3 27B Instruct model, prepared by the MLX Community for efficient inference on Apple Silicon hardware. The conversion was performed using mlx-vlm v0.1.18.

Overview

Gemma 3 27B IT is Google's instruction-tuned large language model built on the Gemma 3 architecture, featuring 27 billion parameters. This variant has been quantized to 4-bit precision and reformatted for the MLX framework, Apple's machine learning library designed to take full advantage of the unified memory and GPU capabilities of M-series chips.

Multimodal Capabilities

Classified under the image-text-to-text pipeline, this model supports vision-language tasks — accepting both images and text as input and generating text responses. This makes it well suited for:

Image captioning and description
Visual question answering
Document and diagram interpretation
General instruction-following and conversational AI

Why 4-Bit Quantization?

The 4-bit quantization dramatically reduces the memory footprint of the full 27B-parameter model, making it feasible to run locally on consumer Apple hardware such as MacBook Pros and Mac Studios with sufficient unified memory. Despite the reduced precision, 4-bit variants of large models typically retain strong performance across a wide range of tasks.

Provenance

Base model: google/gemma-3-27b-pt
Instruction-tuned source: google/gemma-3-27b-it
Conversion tool: mlx-vlm
License: Gemma (requires agreement to Google's usage license on Hugging Face)

This model is a strong choice for developers and researchers seeking a powerful, multimodal instruction-tuned model that runs natively and efficiently on Apple Silicon without relying on cloud infrastructure.

Explore all local models