Model Details
Qwen3 32B
Quick answer: Qwen3 32B is Alibaba's April 2025 dense 32B model scoring 93.8% Arena Hard, 70.3% BFCL, 65.7% LiveCodeBench, and 50.2% Aider. Apache 2.0 with both thinking (reasoning) and non-thinking modes.
At a Glance
Where Qwen3 32B leads
- 93.8% Arena Hard — near-frontier instruction quality at 32B dense
- 65.7% LiveCodeBench — excellent coding
- 70.3% BFCL — strong function calling
- 50.2% Aider — solid coding agent performance
- Apache 2.0 — fully open
- Supports both thinking and non-thinking inference modes
Where it lags
- 16.3% AetherCode — limited on this agentic benchmark
- Dense 32B: higher per-token cost than MoE equivalent (Qwen3 30B A3B)
- April 2025 release
Best for: General reasoning and coding; teams wanting a single dense model without MoE complexity; balanced tool-calling and chat pipelines.
What Qwen3 32B Is
Qwen3 32B is the dense 32B model in Alibaba's Qwen3 generation, released April 2025 alongside the MoE variants. Unlike Qwen3 30B A3B (MoE, 3B active), the 32B dense model activates all parameters at inference — trading higher per-token compute for simpler deployment and no MoE routing overhead.
The hybrid thinking/non-thinking mode support means Qwen3 32B can operate in chain-of-thought reasoning mode for complex tasks or fast non-thinking mode for simple responses — controlled by a mode parameter or special tokens.
The 93.8% Arena Hard is among the highest scores for an open-weight dense 32B model, comparable to the MoE-based Qwen3 30B A3B (91%).
Specifications
| Field | Value |
|---|---|
| Organization | Alibaba |
| License | Apache 2.0 |
| HuggingFace | Qwen/Qwen3-32B |
| Release date | April 29, 2025 |
| Parameters | 32B (dense) |
| Modality | Text only |
| Context window | 128K tokens |
| Inference modes | Thinking / Non-thinking |
Pricing
Open weights under Apache 2.0 — self-host. Also available via Alibaba Cloud and major providers.
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| Arena Hard | 93.8% | Benchgen evaluation | 2025-04 |
| BFCL | 70.3% | Benchgen evaluation | 2025-04 |
| LiveCodeBench | 65.7% | Benchgen evaluation | 2025-04 |
| Aider | 50.2% | Benchgen evaluation | 2025-04 |
| AetherCode | 16.3% | Benchgen evaluation | 2025-04 |
Qwen3 32B vs Alternatives
| Model | Arena Hard | LiveCodeBench | Params | License |
|---|---|---|---|---|
| Qwen3 32B | 93.8% | 65.7% | 32B dense | Apache 2.0 |
| Qwen3 30B A3B | 91% | 62.6% | 30B/3B active | Apache 2.0 |
| Qwen3 235B A22B | 95.6% | 65.9% | 235B/22B active | Apache 2.0 |
| QwQ-32B | — | — | 32B | Apache 2.0 |
Qwen3 32B vs 30B A3B: 93.8% vs 91% Arena Hard at 10.7x active param cost. Use 30B A3B for throughput; 32B for simpler deployment or when MoE is not supported.
Related Models & Benchmarks
- Compare: Qwen3 30B A3B · Qwen3 235B A22B · QwQ-32B
- Benchmarks: Arena Hard · LiveCodeBench · BFCL
Frequently Asked Questions
What is Qwen3 32B?
Alibaba's April 2025 dense 32B model scoring 93.8% Arena Hard, 65.7% LiveCodeBench, 70.3% BFCL. Apache 2.0 with thinking and non-thinking modes.What is the difference between Qwen3 32B and Qwen3 30B A3B?
Qwen3 32B is a dense model (all 32B params active). Qwen3 30B A3B is MoE (only 3B active). Qwen3 32B has higher Arena Hard (93.8% vs 91%) but higher per-token cost. MoE is preferred for high-throughput deployments.Specs from Alibaba's Qwen3 32B release (April 2025) and Benchgen evaluations. Last updated 2026-07-24.
Benchmark Leaderboards
This model isn’t on any benchmark leaderboard yet.