Model Details
Qwen3 VL 4B Thinking
Quick answer: Qwen3 VL 4B Thinking is Alibaba's February 2026 ultra-compact vision-language model scoring 73.6% MMLU-Pro, 67.3% BFCL-v3, and 36.8% Arena Hard v2. Apache 2.0 — 4B params.
At a Glance
Where Qwen3 VL 4B Thinking leads
- 73.6% MMLU-Pro — excellent for 4B multimodal
- 67.3% BFCL-v3 — best BFCL-v3 in the Qwen3 VL family
- Apache 2.0 — fully open
- 4B: minimal inference cost
Where it lags
- 36.8% Arena Hard — weakest instruction quality in Qwen3 VL family
- Smallest tier: limited complex reasoning
Best for: Ultra-compact multimodal deployments; BFCL function calling at minimal inference cost; edge/embedded vision-language tasks.
What Qwen3 VL 4B Thinking Is
Qwen3 VL 4B Thinking is the smallest model in Alibaba's Qwen3 VL thinking family. Notably, it has the highest BFCL-v3 (67.3%) among the three Qwen3 VL variants, suggesting specialised function-calling training effectiveness at 4B scale.
Specifications
| Field | Value |
|---|---|
| Organization | Alibaba |
| License | Apache 2.0 |
| HuggingFace | Qwen/Qwen3-VL-4B-Thinking |
| Release date | February 2026 |
| Parameters | 4B |
| Modality | Text and vision |
Pricing
Open weights under Apache 2.0 — self-host at no cost.
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| MMLU-Pro | 73.6% | Benchgen evaluation | 2026-02 |
| BFCL-v3 | 67.3% | Benchgen evaluation | 2026-02 |
| Arena Hard v2 | 36.8% | Benchgen evaluation | 2026-02 |
Qwen3 VL 4B vs Alternatives
| Model | MMLU-Pro | BFCL-v3 | Arena Hard | Size |
|---|---|---|---|---|
| Qwen3 VL 4B Thinking | 73.6% | 67.3% | 36.8% | 4B |
| Qwen3 VL 8B Thinking | 77.3% | 63.0% | 51.1% | 8B |
| Qwen3 VL 32B Thinking | 82.1% | 71.7% | 60.5% | 32B |
Qwen3 VL 4B has highest BFCL-v3 (67.3%) vs the 8B (63.0%), despite being smaller — good choice for function-calling at minimum compute.
Related Models & Benchmarks
- Compare: Qwen3 VL 8B Thinking · Qwen3 VL 32B Thinking
- Benchmarks: MMLU-Pro · BFCL-v3 · Arena Hard v2
Frequently Asked Questions
What is Qwen3 VL 4B Thinking?
Alibaba's February 2026 4B vision-language model scoring 73.6% MMLU-Pro, 67.3% BFCL-v3. Apache 2.0 — smallest Qwen3 VL, best BFCL-v3 in family.Specs from Alibaba's Qwen3 VL 4B Thinking release (February 2026) and Benchgen evaluations. Last updated 2026-07-24.
Benchmark Leaderboards
This model isn’t on any benchmark leaderboard yet.