Qwen3-VL-4B-Instruct
- In Input modalities: TextImage
- Reasoning
- Tool use
- Open weights
Prices
0 offerings · prices as listedNo priced offering is listed for this model yet.
Details
- Licence
- apache-2.0
- Family
- qwen-vl
Not stated on Alibaba Qwen’s pages: max input, output modalities, reasoning efforts, reasoning can be off, max thinking budget, structured output, built-in tools, knowledge cutoff, aliases.
Run locally
Weights: Qwen/Qwen3-VL-4B-Instruct- Parameters 4.4B
- Dense
- Native precision BF16
- Layers 36
- KV cache per token 144 KB
- Config positions 256K (config.json max_position_embeddings)
- API context 250K (maker’s API)
- Needs
- ≈ 9.9 GB (weights 8.3 (maker’s file) + KV cache 1.1 + overhead 0.5)
- Usable
- 21.6 GB (90% of its memory)
- Generation speed
- 50–75 tok/s
- Longest context
- 91K (longer doesn’t fit)
- First token
- about 0.14–0.21 s for a 2K prompt
| Precision | Weights | Needs at 8K | On your hardware | Longest context |
|---|---|---|---|---|
| BF16 (native)maker’s file | 8.3 GB | ≈ 9.9 | ●Runs well 50–75 tok/s | 91K |
| Q8_0maker’s file | 4.0 GB | ≈ 5.6 | ●Runs well 92–138 tok/s | 122K |
| FP8maker’s file | 5.6 GB | ≈ 7.2 | ●Runs well 70–105 tok/s | 110K |
| Q6_K | 3.4 GB | ≈ 5.0 | ●Runs well 104–156 tok/s | 126K |
| Q5_K_M | 2.9 GB | ≈ 4.6 | ●Runs well 115–173 tok/s | 129K |
| Q5_0 | 2.8 GB | ≈ 4.5 | ●Runs well 118–178 tok/s | 129K |
| Q4_K_Mmaker’s file | 2.3 GB | ≈ 4.0 | ●Runs well 136–204 tok/s | 133K |
| Q4_0 | 2.3 GB | ≈ 3.9 | ●Runs well 136–204 tok/s | 133K |
| IQ4_XS | 2.2 GB | ≈ 3.8 | ●Runs well 141–212 tok/s | 134K |
| Q3_K_Mquality drops | 2.0 GB | ≈ 3.6 | ●Runs well 149–224 tok/s | 135K |
| IQ3_XXSquality drops | 1.6 GB | ≈ 3.2 | ●Runs well 173–260 tok/s | 139K |
| Q2_Kquality drops | 1.5 GB | ≈ 3.2 | ●Runs well 175–263 tok/s | 139K |
| IQ2_XXSquality drops | 1.1 GB | ≈ 2.7 | ●Runs well 214–322 tok/s | 142K |
Click a row to choose a precision. “Maker’s file” sizes are the maker’s own files; others are parameters × bits per weight.
Compared with other hardware
Each at the highest precision it runs. Click one for details.
Smallest that runs it (Q4_K_M, 8K): GPU GeForce RTX 3080; unified memory Apple M5 · 16 GB.
Run commands
vllm serve Qwen/Qwen3-VL-4B-Instruct --max-model-len 8192
Memory to fine-tune
Batch 1, 2,048-token sequences, gradient checkpointing; LoRA and QLoRA train rank-16 adapters on the attention projections.
Estimates; facts from the maker’s Hugging Face repositories, as of 2026-11-06. How it’s estimated
Cited scores
From third-party leaderboards, not measured by usThis model isn’t on any leaderboard we cite yet.
Changes
All changes →No changes recorded yet.
Sources
- Qwen3-VL-4B-Instruct (opens in a new tab) huggingface.co · checked Nov 5, 2026
- Qwen/Qwen3-VL-4B-Instruct config.json (opens in a new tab) huggingface.co · checked Nov 3, 2026
- Qwen/Qwen3-VL-4B-Instruct model card (opens in a new tab) huggingface.co · checked Nov 2, 2026
- Qwen/Qwen3-VL-4B-Instruct-FP8 files (opens in a new tab) huggingface.co · checked Oct 31, 2026
- Qwen/Qwen3-VL-4B-Instruct files (opens in a new tab) huggingface.co · checked Nov 2, 2026
- Qwen/Qwen3-VL-4B-Instruct-GGUF files (opens in a new tab) huggingface.co · checked Nov 2, 2026
Alibaba Qwen’s official policies: data use, privacy, terms →
Some values on this page are still waiting to be checked against official pages: a filled dot is verified, a hollow one is not yet (with or without a source on file). How we verify data
Alibaba Qwen · China · qwen.ai
Related models
Snapshots, the same family and other models from this provider