Qwen3 family: which size to run
5 sizes, 1.7B–30.5B parameters · Alibaba
Sizes
| Model | Architecture | |||||||
|---|---|---|---|---|---|---|---|---|
| Qwen3 1.7B | 1.7B | Dense | 32,768 | No LiveBench result for this model in release 2026-06-25 | 10 | 4 | 1.1 | 1 |
| Qwen3 4B Instruct 2507 | 4B | Dense | 262,144 | No LiveBench result for this model in release 2026-06-25 | 11 | 5 | 2.5 | 2.3 |
| Qwen3 8B | 8.2B | Dense | 32,768 | No LiveBench result for this model in release 2026-06-25 | 15 | 9 | 5 | 4.6 |
| Qwen3 14B | 14.8B | Dense | 32,768 | No LiveBench result for this model in release 2026-06-25 | 20 | 14 | 9 | 8.3 |
| Qwen3 30B A3B Instruct 2507 | 30.5B | MoE · 3.3B active | 262,144 | No LiveBench result for this model in release 2026-06-25 | 31 | 25 | 18.6 | 17.2 |
Each size is its own model, with its own scores and memory needs. Q4 estimates are planning figures, not measurements; file sizes are published. The Mac or PC column is for a 32 GB or larger machine (8 GB reserve); machines under 32 GB reserve 25% of their memory (2 to 8 GB), so they need less. The publisher lists 0.6B, 1.7B, 4B, 8B, 14B, 32B (dense); 30B-A3B, 235B-A22B (MoE).
The 4B (August 2025) and 30B A3B (July 2025) here are the “2507” updates (Instruct, non-thinking) of those two sizes; 1.7B, 8B and 14B are the original April 2025 release. Each is its own checkpoint with its own scores.
How to read this table
Q4 means compressed to about 4 bits per weight. GGUF (llama.cpp, Ollama, LM Studio) and MLX (Apple silicon) are file formats; the sizes shown are the published 4-bit files. A dash means no value is published or collected for that size; the reason is in the cell's accessible text.
Which size fits my machine
The largest size whose Q4 planning estimate fits what each machine offers. A planning estimate, not a measurement and not a speed.
A dash means no size of this family fits that machine. The requirement is the size's Q4 estimate with that machine's own reserve: a graphics card reserves 2 GB, as in the sizes table; a Mac or PC reserves 25% of its memory (2 to 8 GB), so below 32 GB its figure is lower than the sizes table's. A server profile reserves 2 GB per graphics card, so its figure is higher.
Smallest machine for each size
The profile with the least memory that still has enough for the size's Q4 estimate, from the hardware catalog.
| Size | Smallest Mac or PC-memory profile | Smallest graphics-card profile |
|---|---|---|
| Qwen3 1.7B | Mac mini 2020 · M1 · 8 GB | GTX 1650 · 4 GB |
| Qwen3 4B Instruct 2507 | Mac mini 2020 · M1 · 8 GB | RTX 3050 · 6 GB |
| Qwen3 8B | Mac mini 2020 · M1 · 16 GB | RTX 3060 · 12 GB |
| Qwen3 14B | Mac mini 2020 · M1 · 16 GB | RTX 4060 Ti · 16 GB |
| Qwen3 30B A3B Instruct 2507 | Mac mini 2024 · M4 · 32 GB | RTX 5090 · 32 GB |
A dash means no profile of that kind in the catalog has enough memory for the size.
Memory by file precision
The same planning formula at three precisions (0.625, 1.125 and 2 bytes per parameter, plus working room and the system reserve). Estimates, not measurements.
| Size | ||||||
|---|---|---|---|---|---|---|
| Qwen3 1.7B | 10 | 4 | 11 | 5 | 13 | 7 |
| Qwen3 4B Instruct 2507 | 11 | 5 | 14 | 8 | 18 | 12 |
| Qwen3 8B | 15 | 9 | 20 | 14 | 28 | 22 |
| Qwen3 14B | 20 | 14 | 28 | 22 | 44 | 38 |
| Qwen3 30B A3B Instruct 2507 | 31 | 25 | 50 | 44 | 82 | 76 |
Family tree
Family, sizes and, for each size, the two 4-bit files the tables use (GGUF Q4_K_M and MLX 4-bit) when one was collected. Other precisions are not drawn here.
The same tree as a list:
- Qwen3 · Alibaba
- 1.7B
- GGUF Q4_K_M · 1.1 GB · unsloth/Qwen3-1.7B-GGUF, read Oct 23, 2026
- MLX 4-bit · 1 GB · lmstudio-community/Qwen3-1.7B-MLX-4bit, read Oct 23, 2026
- 4B Instruct 2507
- GGUF Q4_K_M · 2.5 GB · unsloth/Qwen3-4B-Instruct-2507-GGUF, read Oct 23, 2026
- MLX 4-bit · 2.3 GB · mlx-community/Qwen3-4B-Instruct-2507-4bit, read Oct 23, 2026
- 8B
- GGUF Q4_K_M · 5 GB · Qwen/Qwen3-8B-GGUF, read Oct 23, 2026
- MLX 4-bit · 4.6 GB · mlx-community/Qwen3-8B-4bit, read Oct 23, 2026
- 14B
- GGUF Q4_K_M · 9 GB · Qwen/Qwen3-14B-GGUF, read Oct 23, 2026
- MLX 4-bit · 8.3 GB · lmstudio-community/Qwen3-14B-MLX-4bit, read Oct 23, 2026
- 30B A3B Instruct 2507 · 3.3B active
- GGUF Q4_K_M · 18.6 GB · unsloth/Qwen3-30B-A3B-Instruct-2507-GGUF, read Oct 23, 2026
- MLX 4-bit · 17.2 GB · lmstudio-community/Qwen3-30B-A3B-Instruct-2507-MLX-4bit, read Oct 23, 2026
- 1.7B
Learn more
See Qwen3 1.7B
The smallest size of the family: its scores, memory estimate and the machines it runs on.
Sources
Where the family and its sizes are named, and where each number comes from.
- QwenLM/Qwen3 README · “Dense and Mixture-of-Experts (MoE) models of various sizes, available in 0.6B, 1.7B, 4B, 8B, 14B, 32B and 30B-A3B, 235B-A22B.” · read Oct 17, 2026
- QwenLM/Qwen3 README · “we are excited to introduce the updated Qwen3-2507 in two variants, Qwen3-Instruct-2507 and Qwen3-Thinking-2507, and three sizes, 235B-A22B, 30B-A3B, and 4B.” · read Oct 17, 2026
- Qwen/Qwen3-Coder-30B-A3B-Instruct model card · “Qwen3-Coder is available in multiple sizes. Today, we're excited to introduce Qwen3-Coder-30B-A3B-Instruct.” · read Oct 17, 2026
- deepseek-ai/DeepSeek-R1-0528-Qwen3-8B model card · “Meanwhile, we distilled the chain-of-thought from DeepSeek-R1-0528 to post-train Qwen3 8B Base, obtaining DeepSeek-R1-0528-Qwen3-8B. … The model architecture of DeepSeek-R1-0528-Qwen3-8B is identical to that of Qwen3-8B, but it shares the same tokenizer configuration as DeepSeek-R1-0528.” · read Oct 17, 2026
- Qwen3 1.7B model card
- Qwen3 4B Instruct 2507 model card
- Qwen3 8B model card
- Qwen3 14B model card
- Qwen3 30B A3B Instruct 2507 model card
- LiveBench 2026-06-25 · the only quality source on this page; a dash means the release does not include the model · snapshot collected Sep 26, 2026
- Published file sizes: Hugging Face repositories, read Oct 23, 2026
- Methodology · how the Q4 planning estimate is computed · Learn