Leaderboard / Alibaba (Qwen)

Qwen3 4B Thinking 2507

Name: Qwen3 4B Thinking 2507
Author: Alibaba (Qwen)

Alibaba (Qwen) · Qwen3 · released 2025-08-06 · data checked 2026-11-01 Needs verification

Qwen3 4B Thinking 2507 is a 4B-parameter dense model from Alibaba (Qwen) (China), released 2025-08-06 under the Apache-2.0 license. It scores 55.7 on the Small LM Index, #8 of 21 tiny models. It needs about 4.2 GB of memory at 4-bit with an 8K-token context.

Small LM Index

55.7

#8 of 21 scored tiny models

Likely 51.2–59.9 · from 4 of 6 core benchmarks

Parameters

Dense

Price per 1M tokens

—

No hosted API on OpenRouter; self-host

Context window

262K

tokens

Benchmarks

Each bar shows this model's score. The tick marks the best tiny model on that benchmark. Sources: 4 reported by Alibaba (Qwen). How we label sources.

MMLU-Pro Lab-reported Knowledge74.0

Class best: Qwen3.5 4B 79.1

GPQA Diamond Lab-reported Knowledge65.8

Class best: Nanbeige4.2 3B 87.4

AIME Lab-reported Math · AIME 202581.3

Class best: Spark-X2.6 4B 91.2

LiveCodeBench Lab-reported Coding · LiveCodeBench v655.2

Class best: LFM3 1.7B 73.5

Not reported: estimated

What the index model expects from this model's other scores. Not a measurement.

Humanity's Last Exam≈ 8

SWE-bench Verified≈ 39

Specifications

Lab: Alibaba (Qwen) (CN · China)
Architecture: Dense transformer
Total parameters: 4B
Active parameters: All (dense)
Size class: Tiny
Context window: 262K tokens
Input modalities: text
Reasoning: Reasoning (thinking)
License: Apache-2.0
Openness: Open weights
Memory to run: ~4.2 GB at 4-bit, 8K context (more sizes)
Hugging Face: 440.9K downloads (30 days) · 629 likes

Weights on Hugging Face ↗

Something wrong here? Report it.

Run it locally

Uses unsloth/Qwen3-4B-Thinking-2507-GGUF (Q4_K_M). Swap the tag for another size: F16, IQ4_NL, IQ4_XS, Q2_K, Q2_K_L, Q3_K_M, Q3_K_S, Q4_0….

Ollama

ollama run hf.co/unsloth/Qwen3-4B-Thinking-2507-GGUF:Q4_K_M

llama.cpp

llama-server -hf unsloth/Qwen3-4B-Thinking-2507-GGUF:Q4_K_M

Then open http://localhost:8080 for a chat window.

MLX (Mac)

pip install -U mlx-lm
mlx_lm.chat --model mlx-community/Qwen3-4B-Thinking-2507-4bit

vLLM (GPU server)

vllm serve Qwen/Qwen3-4B-Thinking-2507

Full-precision weights, for a server GPU.

Open in LM Studio

Memory needed: weights, context cache and runtime buffers.

Quantization	4K context	32K context	256K context
16-bit F16	9.2 GB	13 GB	47 GB
8-bit Q8_0	5.4 GB	9.7 GB	43 GB
4-bit Q4_K_M	3.6 GB	7.8 GB	42 GB

Check it against your GPU, Mac or phone →

Qwen3 4B Thinking 2507

Benchmarks

Not reported: estimated

Specifications

Compare with

Ranked in

Sources

Run it locally