Leaderboard / Alibaba (Qwen)

Qwen3 30B-A3B Thinking 2507

Name: Qwen3 30B-A3B Thinking 2507
Author: Alibaba (Qwen)

Alibaba (Qwen) · Qwen3 · released 2025-07-30 · data checked 2026-11-01

Qwen3 30B-A3B Thinking 2507 is a 30.5B-parameter Mixture-of-Experts model (3.3B active per token) from Alibaba (Qwen) (China), released 2025-07-30 under the Apache-2.0 license. It scores 59.3 on the Small LM Index, #21 of 34 medium models. It needs about 20.4 GB of memory at 4-bit with an 8K-token context.

Small LM Index

59.3

#21 of 34 scored medium models

Likely 55.4–62.9 · from 6 of 6 core benchmarks

Parameters

30.5B

3.3B active · MoE

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 medium model on that benchmark. Sources: 6 reported by Alibaba (Qwen), 1 measured by another lab. How we label sources.

MMLU-Pro Lab-reported Knowledge80.9

Class best: Qwen3.6 27B 86.2

GPQA Diamond Lab-reported Knowledge73.4

Class best: Nanbeige4.3 33B 89.5

Humanity's Last Exam Lab-reported Reasoning9.8

Class best: Qwen3.8 27B 30.8

AIME Lab-reported Math · AIME 202585.0

Class best: Ternary Bonsai 2 27B 95.8

LiveCodeBench Lab-reported Coding · LiveCodeBench v666.0

Class best: Qwen3.8 27B 90.3

SWE-bench Verified Other lab Coding · Measured by Z.ai / NVIDIA (OpenHands)22.0

Class best: Qwen3.6 27B 77.2

IFBench Lab-reported Instruction following51.5

Class best: Sage-A1.5 81.6

Specifications

Lab: Alibaba (Qwen) (CN · China)
Architecture: Mixture-of-Experts
Total parameters: 30.5B
Active parameters: 3.3B
Size class: Medium
Context window: 262K tokens
Input modalities: text
Reasoning: Reasoning (thinking)
License: Apache-2.0
Openness: Open weights
Memory to run: ~20.4 GB at 4-bit, 8K context (more sizes)
Hugging Face: 66.6K downloads (30 days) · 391 likes

Weights on Hugging Face ↗

Compare with

Qwen3 30B-A3B Thinking 2507 vs Muse Glimmer 30B (29.6B, index 72.6)
Qwen3 30B-A3B Thinking 2507 vs Gemma 4 31B (30.7B, index 70.1)
Qwen3 30B-A3B Thinking 2507 vs Granite 4.2 30B (30B, index 67.0)
Qwen3 30B-A3B Thinking 2507 vs GLM-4.7-Flash (31B, index 66.1)
Qwen3 30B-A3B Thinking 2507 vs Kanana-2 30B-A3B Thinking (30.7B, index 50.7)

Sources

Something wrong here? Report it.

Run it locally

Uses unsloth/Qwen3-30B-A3B-Thinking-2507-GGUF (Q4_K_M). Swap the tag for another size: BF16, IQ4_NL, IQ4_XS, Q2_K, Q2_K_L, Q3_K_M, Q3_K_S, Q4_0….

Ollama

ollama run hf.co/unsloth/Qwen3-30B-A3B-Thinking-2507-GGUF:Q4_K_M

llama.cpp

llama-server -hf unsloth/Qwen3-30B-A3B-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/Qwen3-30B-A3B-Thinking-2507-4bit

vLLM (GPU server)

vllm serve Qwen/Qwen3-30B-A3B-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	63 GB	65 GB	88 GB
8-bit Q8_0	34 GB	37 GB	59 GB
4-bit Q4_K_M	20 GB	23 GB	45 GB

Check it against your GPU, Mac or phone →