Qwen/Qwen3-4B-Instruct-2507

Name: Qwen/Qwen3-4B-Instruct-2507
Author: Alibaba

Run locally on Apple devices with Mirai

Type: local
From: Alibaba
Quantization: No
Parameters: 4B
Size: 7.5 GB
Source

Qwen3-4B-Instruct-2507 is an updated instruct-tuned language model from the Qwen team, delivering a significant leap in capability over its predecessor within the same compact 4-billion-parameter footprint. It operates exclusively in non-thinking mode — generating direct, efficient responses without intermediate reasoning blocks.

Qwen3-4B-Instruct-2507 benchmark performance

Architecture & Specs

Parameters: 4.0B total (3.6B non-embedding)
Architecture: Causal language model with 36 layers, GQA (32 query heads, 8 KV heads)
Context Length: 262,144 tokens natively
Training: Full pretraining followed by post-training alignment

Key Improvements

This release brings substantial upgrades across the board compared to the original Qwen3-4B non-thinking variant:

Reasoning & Math: Dramatic jumps on competition-level benchmarks — AIME25 scores rise from 19.1 to 47.4, and ZebraLogic from 35.2 to 80.2.
Knowledge: Near-parity with much larger models on MMLU-Pro (69.6) and GPQA (62.0), rivaling Qwen3-30B-A3B.
Coding: Improved performance on LiveCodeBench and MultiPL-E, outperforming GPT-4.1-nano across most coding tasks.
Alignment & Creative Writing: Markedly stronger on subjective tasks — Arena-Hard v2 jumps from 9.5 to 43.4, and WritingBench from 68.5 to 83.4.
Agentic Use: Strong tool-calling results on BFCL-v3 and TAU benchmarks, with native support for MCP-based agent workflows via Qwen-Agent.
Multilingual Coverage: Enhanced long-tail knowledge across multiple languages, with notable gains on PolyMATH (31.1, up from 16.6).

Use Cases

Ideal for resource-constrained deployments needing strong general-purpose performance — instruction following, code generation, long-document understanding, multilingual tasks, and agentic tool use. Compatible with HuggingFace Transformers, vLLM, SGLang, Ollama, LMStudio, and llama.cpp.

1	import Uzu
2
3	public func runChat() async throws {
4	let engineConfig = EngineConfig.create()
5	let engine = try await Engine.create(config: engineConfig)
6
7	guard let model = try await engine.model(identifier: "Qwen/Qwen3-4B-Instruct-2507") else {
8	return
9	}
10	for try await update in try await engine.download(model: model).iterator() {
11	print("Download progress: \(update.progress())")
12	}
13
14	let messages = [
15	ChatMessage.system().withText(text: "You are a helpful assistant"),
16	ChatMessage.user().withText(text: "Tell me a short, funny story about a robot")
17	]
18	let session = try await engine.chat(model: model, config: .create())
19	let stream = await session.replyWithStream(input: messages, config: .create())
20	var message: ChatMessage? = nil
21	for try await update in stream.iterator() {
22	switch update {
23	case .replies(let replies):
24	message = replies.last?.message
25	case .error(let error):
26	print("Error: \(error)")
27	}
28	}
29	print("Text: \(message?.text() ?? "empty")")
30	}

Architecture & Specs

Key Improvements

Use Cases

Other local models from Alibaba