Quick answer: QwQ 32B is Alibaba's March 2025 open-weight reasoning model — Qwen's answer to DeepSeek-R1 in the open reasoning space. It scores 66.4% on BFCL v3. Under Apache 2.0, it is freely available for commercial use. QwQ 32B uses extended chain-of-thought reasoning, similar to the o-series, to improve accuracy on complex tasks.
Where QwQ 32B leads
Where it lags
Best for: Open-weight deployments requiring reasoning capability without o-series proprietary pricing; agentic function-calling tasks; research into open reasoning models.
QwQ 32B (released March 5, 2025) is Alibaba's open reasoning model — part of the Qwen family but specifically designed for extended chain-of-thought reasoning, similar in architecture to OpenAI's o1 and DeepSeek's R1. The "QwQ" name reflects the model's exploratory, questioning reasoning approach.
The model's 66.4% BFCL v3 score demonstrates strong function-calling capability when combined with reasoning — notably higher than standard 32B models without thinking modes. This makes QwQ 32B particularly suitable for tool-using agentic applications where accuracy on function calls matters.
QwQ 32B fills the niche of "open-weight reasoning model at 32B scale" — sitting between Qwen3 32B (better general performance, faster, hybrid thinking) and DeepSeek-R1 (larger scale, stronger overall reasoning). For teams that need Apache 2.0 licensing with a reasoning model at 32B scale, QwQ 32B is the primary option.
| Field | Value |
|---|---|
| Organization | Alibaba / Qwen |
| Parameters | 32B (dense) |
| Context window | 131,072 tokens |
| License | Apache 2.0 |
| HuggingFace | Qwen/QwQ-32B |
| Release date | March 5, 2025 |
| Knowledge cutoff | September 2024 |
| Modality | Text only |
| Architecture | Extended chain-of-thought reasoning |
QwQ 32B is available as open weights under Apache 2.0 — free to self-host. Available via Alibaba Cloud DashScope and third-party providers.
| Benchmark | Score | Source | Date |
|---|---|---|---|
| BFCL v3 | 66.4% | Benchgen evaluation | 2025-07 |
| Model | BFCL v3 | Reasoning | License | Params |
|---|---|---|---|---|
| QwQ 32B | 66.4% | Yes (CoT) | Apache 2.0 | 32B |
| Kimi K2 Instruct | 59.1% | No | Apache 2.0 | — |
| Qwen3 32B | — | Hybrid | Apache 2.0 | 32B |
| o4-mini | — | Yes (o-series) | Proprietary | — |
QwQ 32B vs Kimi K2 Instruct: higher BFCL v3 (66.4% vs 59.1%) with reasoning enabled. QwQ 32B vs Qwen3 32B: QwQ is a dedicated reasoning model with chain-of-thought; Qwen3 32B has hybrid thinking that can be enabled/disabled. For purely function-calling tasks, QwQ 32B shows stronger results.
Specs from Qwen's official QwQ 32B release (March 2025) and Benchgen evaluations. Last updated 2026-07-24.
This model isn’t on any benchmark leaderboard yet.