Model Details
Qwen3 30B A3B
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 | 91% | Benchgen evaluation | 2025-05 |
| LiveCodeBench | 62.6% | Benchgen evaluation | 2025-05 |
| BFCL | 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 | 95.6% | 65.9% | 22B | Apache 2.0 |
| QwQ-32B | — | — | 32B | Apache 2.0 |
| 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 · QwQ-32B · Llama 3.3 70B Instruct
- Benchmarks: Arena Hard · LiveCodeBench · BFCL
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.
Benchmark Leaderboards
This model isn’t on any benchmark leaderboard yet.