# Qwen3 30B A3B ![Organization](https://img.shields.io/badge/Org-Alibaba-ff6a00) ![License](https://img.shields.io/badge/License-Apache%202.0-22c55e) ![Released](https://img.shields.io/badge/Released-May%202025-555) ![MoE](https://img.shields.io/badge/Type-MoE-6366f1) > **Quick answer:** Qwen3 30B A3B is Alibaba's May 2025 compact MoE model with 30B total parameters (3B active), scoring 91% Arena Hard, 62.6% LiveCodeBench, and 69.1% BFCL. Apache 2.0 — offers near-frontier Arena Hard at minimal inference cost. ## At a Glance **Where Qwen3 30B A3B leads** - 91% Arena Hard — near-frontier instruction quality at only 3B active params - 62.6% LiveCodeBench — excellent coding for its compute cost - 69.1% BFCL — strong function calling - MoE efficiency: 3B active params → low inference cost - Apache 2.0 **Where it lags** - 91% Arena Hard vs 95.6% for Qwen3 235B A22B — significant gap - Less capacity for complex reasoning than larger models - May 2025 release: pre-dates some newer models **Best for:** Cost-efficient deployment requiring near-frontier instruction quality; API servers with limited GPU; teams wanting Qwen3-class performance at 3B active-param inference cost. ## What Qwen3 30B A3B Is Qwen3 30B A3B is the compact MoE model in Alibaba's Qwen3 generation. "30B A3B" means 30 billion total parameters with only 3 billion active per token inference pass. At 3B active params, inference cost is similar to a compact dense 3B model, while the full 30B capacity allows much higher quality. The 91% Arena Hard score with 3B active parameters is remarkable — this is competitive with Qwen3 32B dense variants and near-frontier proprietary models, at a fraction of the per-token compute cost. ## Specifications | Field | Value | |---|---| | Organization | Alibaba | | License | Apache 2.0 | | HuggingFace | Qwen/Qwen3-30B-A3B | | Release date | May 28, 2025 | | Parameters | 30B total / 3B active (MoE) | | Modality | Text only | | Context window | 128K tokens | ## Pricing Open weights under Apache 2.0 — self-host. Also available via Alibaba Cloud and third-party providers. ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [Arena Hard](/benchmarks/arena-hard-v2) | 91% | Benchgen evaluation | 2025-05 | | [LiveCodeBench](/benchmarks/livecodebench) | 62.6% | Benchgen evaluation | 2025-05 | | [BFCL](/benchmarks/bfcl-v3) | 69.1% | Benchgen evaluation | 2025-05 | ## Qwen3 30B A3B vs Alternatives | Model | Arena Hard | LiveCodeBench | Active Params | License | |---|---|---|---|---| | **Qwen3 30B A3B** | 91% | 62.6% | 3B | Apache 2.0 | | [Qwen3 235B A22B](/models/alibaba/qwen3-235b-a22b) | 95.6% | 65.9% | 22B | Apache 2.0 | | [QwQ-32B](/models/qwen/qwq-32b) | — | — | 32B | Apache 2.0 | | [Llama 3.3 70B Instruct](/models/meta/llama-3-3-70b-instruct) | — | — | 70B | Llama 3.3 | Qwen3 30B A3B vs 235B A22B: 91% vs 95.6% Arena Hard at 7x lower active-param cost. For most production deployments, 30B A3B is the efficiency winner. ## Related Models & Benchmarks - Compare: [Qwen3 235B A22B](/models/alibaba/qwen3-235b-a22b) · [QwQ-32B](/models/qwen/qwq-32b) · [Llama 3.3 70B Instruct](/models/meta/llama-3-3-70b-instruct) - Benchmarks: [Arena Hard](/benchmarks/arena-hard-v2) · [LiveCodeBench](/benchmarks/livecodebench) · [BFCL](/benchmarks/bfcl-v3) ## Frequently Asked Questions

What is Qwen3 30B A3B?

Alibaba's May 2025 compact MoE model: 30B total parameters, 3B active at inference. 91% Arena Hard, 62.6% LiveCodeBench, 69.1% BFCL. Apache 2.0.

How does Qwen3 30B A3B compare to Qwen3 235B A22B?

235B A22B leads on Arena Hard (95.6% vs 91%) and LiveCodeBench (65.9% vs 62.6%) with 7.3x more active params (22B vs 3B). Choose 30B A3B for compute-cost efficiency; 235B A22B for maximum capability.

--- *Specs from Alibaba's Qwen3 30B A3B release (May 2025) and Benchgen evaluations. Last updated 2026-07-24.*

Qwen3 30B A3B — Evaluations

Evaluations