Qwen3.5-4B
- In Input modalities: TextImage Out Output modalities: Text
- 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
Not stated on Alibaba Qwen’s pages: max input, reasoning efforts, reasoning can be off, max thinking budget, structured output, built-in tools, knowledge cutoff, aliases.
Run locally
Weights: Qwen/Qwen3.5-4B- Parameters 4.7B
- Dense
- Native precision BF16
- Layers 32
- KV cache per token 32 KB
- Native context 256K (model card)
- Config positions 256K (config.json max_position_embeddings)
- API context 256K (maker’s API)
- Needs
- ≈ 9.4 GB (weights 8.7 (maker’s file) + KV cache 0.3 + overhead 0.5)
- Usable
- 21.6 GB (90% of its memory)
- Generation speed
- 53–79 tok/s
- Longest context
- 256K (the model’s limit)
- First token
- about 0.15–0.23 s for a 2K prompt
| Precision | Weights | Needs at 8K | On your hardware | Longest context |
|---|---|---|---|---|
| BF16 (native)maker’s file | 8.7 GB | ≈ 9.4 | ●Runs well 53–79 tok/s | 256K |
| Q8_0 | 4.6 GB | ≈ 5.4 | ●Runs well 97–145 tok/s | 256K |
| FP8 | 4.3 GB | ≈ 5.1 | ●Runs well 102–153 tok/s | 256K |
| Q6_K | 3.6 GB | ≈ 4.3 | ●Runs well 123–185 tok/s | 256K |
| Q5_K_M | 3.1 GB | ≈ 3.8 | ●Runs well 141–211 tok/s | 256K |
| Q5_0 | 3.0 GB | ≈ 3.7 | ●Runs well 145–218 tok/s | 256K |
| Q4_K_M | 2.6 GB | ≈ 3.4 | ●Runs well 163–244 tok/s | 256K |
| Q4_0 | 2.4 GB | ≈ 3.2 | ●Runs well 174–262 tok/s | 256K |
| IQ4_XS | 2.3 GB | ≈ 3.1 | ●Runs well 184–276 tok/s | 256K |
| Q3_K_Mquality drops | 2.1 GB | ≈ 2.9 | ●Runs well 198–297 tok/s | 256K |
| IQ3_XXSquality drops | 1.7 GB | ≈ 2.4 | ●Runs well 246–369 tok/s | 256K |
| Q2_Kquality drops | 1.6 GB | ≈ 2.4 | ●Runs well 250–375 tok/s | 256K |
| IQ2_XXSquality drops | 1.1 GB | ≈ 1.9 | ●Runs well 343–515 tok/s | 256K |
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/Qwen3.5-4B --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 us- Epoch AI · OTIS Mock AIME 2024–2025 55.8%
#107 of 159 listed models · reasoning off · ±6.16
- Epoch AI, 'Capabilities & benchmarking'. Published online at epoch.ai. Retrieved from 'https://epoch.ai/benchmarks'. Licensed CC BY 4.0. · License · as of Nov 4, 2026
- changes we made
Changes
All changes →No changes recorded yet.
Sources
- Qwen3.5-4B (opens in a new tab) huggingface.co · checked Nov 3, 2026
- Qwen/Qwen3.5-4B config.json (opens in a new tab) huggingface.co · checked Oct 31, 2026
- Qwen/Qwen3.5-4B model card (opens in a new tab) huggingface.co · checked Nov 2, 2026
- Qwen/Qwen3.5-4B files (opens in a new tab) huggingface.co · checked Nov 3, 2026
- models.dev: alibaba/qwen3.5-4b (community data, MIT) (opens in a new tab) github.com · checked Oct 22, 2026
- Qwen/Qwen3.5-4B at main (opens in a new tab) huggingface.co · checked Nov 3, 2026
- Qwen/Qwen3.5-4B README.md (opens in a new tab) huggingface.co · checked Nov 4, 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