Qwen 3
Qwen/Qwen3-0.6B- Vendor
- Alibaba
- Parameters
- 600M
- Size
- 1.1 GB
$ brew install mirai$ mirai --model Qwen/Qwen3-0.6BBenchmarks
Qwen 3
Apple M4 Max 128GB
0.12 s
lower is better ↓
uzu0.5.14vs
0.31.2vs
llama.cpp0.3.23Qwen 3
Apple M4 Max 128GB
259 t/s
higher is better ↑
uzu0.5.14vs0.31.2vsllama.cpp0.3.23Qwen 3
Apple M4 Max 128GB
1.58 GB
lower is better ↓
uzu0.5.14vs0.31.2vsllama.cpp0.3.23Benchmarked 7 Aug 2026
Integrate with SDK
https://github.com/trymirai/uzu-swift
| 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
Qwen3-0.6B is an ultra-compact causal language model from Alibaba's Qwen team, part of the third-generation Qwen series. Despite its small footprint — just 0.6 billion parameters (0.44B non-embedding) — it inherits several flagship capabilities from the broader Qwen3 family, making it a compelling option for resource-constrained environments.
Key Capabilities
The model's standout feature is its dual-mode reasoning system, allowing seamless switching between a *thinking mode* for complex logical reasoning, math, and code, and a *non-thinking mode* for efficient general-purpose dialogue. Users can toggle between modes at inference time — or even mid-conversation — using simple `/think` and `/no_think` commands, giving fine-grained control over the cost-quality tradeoff.
Beyond reasoning, Qwen3-0.6B offers:
- Agent and tool-calling support, with structured integration for external tools via frameworks like Qwen-Agent and MCP configurations.
- Multilingual coverage across 100+ languages and dialects, including strong instruction-following and translation performance.
- 32,768-token context length, generous for a model of this size.
Architecture
Built on a dense transformer architecture with 28 layers and grouped-query attention (16 Q heads, 8 KV heads), Qwen3-0.6B is derived from the Qwen3-0.6B-Base model through both pretraining and post-training stages. It is compatible with Hugging Face Transformers (v4.51.0+), as well as serving frameworks like vLLM and SGLang, and local tools such as Ollama, LMStudio, and llama.cpp.
Best For
This model is well-suited for edge deployment, on-device inference, and latency-sensitive applications where a full-scale LLM would be impractical. Its thinking/non-thinking toggle makes it especially versatile — able to handle both quick Q&A and multi-step problem solving within a single lightweight package.
Licensed under Apache 2.0.