Qwen3 models
105 Qwen3 models from Qwen on Hugging Face, from 460M to 480.2B parameters, published by Qwen in BF16, with contexts from 8K to 256K tokens. The smallest official model, Qwen3-Embedding-0.6B, needs about 1.3 GB of VRAM at its published precision; the cheapest live fit is RTX 3070 at $0.088/hr.
Part of the Qwen series · Previous generation: Qwen2.5 · Next generation: Qwen3.5
Pick a size
One row per official Qwen3 size: the VRAM it needs at each precision, the cheapest GPU that holds it today, and the KV cache for a 32K-token context.
| Model | Parameters | Native VRAM | FP8 VRAM | INT4 VRAM | Live GPU fit (native) | Est. $/hr | KV cache at 32K |
|---|---|---|---|---|---|---|---|
| Qwen3-Embedding-0.6B | 596M | 1.3 GB | 0.7 GB | 0.3 GB | RTX 3070 | $0.088/hr | 3.5 GB |
| Qwen3-0.6B | 752M | 1.7 GB | 0.8 GB | 0.4 GB | RTX 3070 | $0.088/hr | 3.5 GB |
| Qwen3-ASR-0.6B | 938M | 2.1 GB | 1.0 GB | 0.5 GB | RTX 3070 | $0.088/hr | 3.5 GB |
| Qwen3-1.7B-Base | 1.7B | 3.8 GB | 1.9 GB | 1.0 GB | RTX 3070 | $0.088/hr | 3.5 GB |
| Qwen3-TTS-12Hz-1.7B-CustomVoice | 1.9B | 4.3 GB | 2.1 GB | 1.1 GB | RTX 3070 | $0.088/hr | not published for this architecture |
| Qwen3-1.7B | 2.0B | 4.5 GB | 2.3 GB | 1.1 GB | RTX 3070 | $0.088/hr | 3.5 GB |
| Qwen3-VL-2B-Instruct | 2.1B | 4.8 GB | 2.4 GB | 1.2 GB | RTX 3070 | $0.088/hr | 3.5 GB |
| Qwen3-ASR-1.7B | 2.3B | 5.3 GB | 2.6 GB | 1.3 GB | RTX 3070 | $0.088/hr | 3.5 GB |
| Qwen3-4B | 4.0B | 9.0 GB | 4.5 GB | 2.2 GB | RTX A5000 | $0.176/hr | 4.5 GB |
| Qwen3-VL-4B-Instruct | 4.4B | 9.9 GB | 5.0 GB | 2.5 GB | RTX A5000 | $0.176/hr | 4.5 GB |
| Qwen3-Embedding-8B | 7.6B | 16.9 GB | 8.5 GB | 4.2 GB | RTX A5000 | $0.176/hr | 4.5 GB |
| Qwen3-VL-Embedding-8B | 8.1B | 18.2 GB | 9.1 GB | 4.6 GB | RTX A5000 | $0.176/hr | 4.5 GB |
| Qwen3-8B | 8.2B | 18.3 GB | 9.2 GB | 4.6 GB | RTX A5000 | $0.176/hr | 4.5 GB |
| Qwen3-VL-8B-Instruct | 8.8B | 19.6 GB | 9.8 GB | 4.9 GB | RTX A5000 | $0.176/hr | 4.5 GB |
| Qwen3-14B | 14.8B | 33.0 GB | 16.5 GB | 8.3 GB | RTX A6000 | $0.363/hr | 5.0 GB |
| Qwen3-30B-A3B | 30.5B | 68.2 GB | 34.1 GB | 17.1 GB | A100 | $1.31/hr | 3.0 GB |
| Qwen3-VL-30B-A3B-Instruct | 31.1B | 69.4 GB | 34.7 GB | 17.4 GB | A100 | $1.31/hr | 3.0 GB |
| Qwen3-32B | 32.8B | 73.2 GB | 36.6 GB | 18.3 GB | A100 | $1.31/hr | 8.0 GB |
| Qwen3-VL-32B-Instruct | 33.4B | 74.6 GB | 37.3 GB | 18.6 GB | A100 | $1.31/hr | 8.0 GB |
| Qwen3-Coder-Next | 79.7B | 178 GB | 89.0 GB | 44.5 GB | AMD MI300X | $2.63/hr | 0.75 GB |
| Qwen3-Next-80B-A3B-Instruct | 81.3B | 182 GB | 90.9 GB | 45.4 GB | AMD MI300X | $2.63/hr | 0.75 GB |
| Qwen3-235B-A22B | 235.1B | 525 GB | 263 GB | 131 GB | AMD MI300X × 3 | $7.89/hr | 5.9 GB |
| Qwen3-VL-235B-A22B-Instruct | 235.7B | 527 GB | 263 GB | 132 GB | AMD MI300X × 3 | $7.89/hr | 5.9 GB |
| Qwen3-Coder-480B-A35B-Instruct | 480.2B | 1073 GB | 537 GB | 268 GB | AMD MI300X × 6 | $15.77/hr | 7.8 GB |
VRAM is weights times a flat 1.2 overhead; the KV cache is a separate per-model figure at 16-bit, shown where the architecture is published. See the methodology.
Official models (50)
Quantized versions (33)
Fine-tunes and community models (22)
| Model | Parameters | Published as | VRAM needed | Live GPU fit | Est. $/hr |
|---|---|---|---|---|---|
| hima-asr | 460M | BF16 | 1.0 GB | RTX 3070 | $0.088/hr |
| Qwen3-4B-DFlash-b16 | 537M | BF16 | 1.2 GB | RTX 3070 | $0.088/hr |
| OmniVoice | 613M | F32 | 2.7 GB | RTX 3070 | $0.088/hr |
| s1-mini | 752M | BF16 | 1.7 GB | RTX 3070 | $0.088/hr |
| Qwen3-8B-speculator.eagle3 | 1.0B | BF16 | 2.3 GB | RTX 3070 | $0.088/hr |
| Qwen3-8B-DFlash-b16 | 1.0B | BF16 | 2.3 GB | RTX 3070 | $0.088/hr |
| Qwen3-32B-speculator.eagle3 | 1.6B | BF16 | 3.5 GB | RTX 3070 | $0.088/hr |
| Qwen3-1.6B-A0.9B | 1.6B | BF16 | 3.5 GB | RTX 3070 | $0.088/hr |
| Qwen3-1.7B-GLB-3D-Coder | 1.7B | F16 | 3.8 GB | RTX 3070 | $0.088/hr |
| Confucius4-R2T2 | 2.0B | BF16 | 4.6 GB | RTX 3070 | $0.088/hr |
| Oc-e-Non-v1 | 4.0B | BF16 | 9.0 GB | RTX A5000 | $0.176/hr |
| LightNav-0 | 4.8B | BF16 | 10.8 GB | RTX A5000 | $0.176/hr |
| avibe | 7.9B | BF16 | 17.7 GB | RTX A5000 | $0.176/hr |
| T-lite-it-2.1 | 8.2B | BF16 | 18.3 GB | RTX A5000 | $0.176/hr |
| Goedel-Prover-V2-8B | 8.2B | BF16 | 18.3 GB | RTX A5000 | $0.176/hr |
| Qwen3-8B-ZeroZero-Heretic | 8.2B | BF16 | 18.3 GB | RTX A5000 | $0.176/hr |
| qwen3vl-resume-parser | 8.8B | BF16 | 19.6 GB | RTX A5000 | $0.176/hr |
| ContextPilot-14B | 14.8B | BF16 | 33.0 GB | RTX A6000 | $0.363/hr |
| lynx-instruct-30b | 30.5B | BF16 | 68.2 GB | A100 | $1.31/hr |
| Qwen3-30B-A3B-abliterated | 30.5B | F32 | 136 GB | AMD MI300X | $2.63/hr |
| Goedel-Prover-V2-32B | 32.8B | BF16 | 73.2 GB | A100 | $1.31/hr |
| Karnak-40B-v1.0 | 40.7B | BF16 | 90.9 GB | RTX PRO 6000 | $1.38/hr |
VRAM is for the precision the model is published in: the weights times a flat 1.2 overhead, with the KV cache not included. See the methodology. The fit shown is the lowest-priced single GPU type that holds the model at that precision, or the lowest-priced multi-GPU set (up to 8) when none does.
Run and fine-tune Qwen3
- How to run Qwen3: engine support, chat template and context flags, with a serving command sized per model.
- How to fine-tune Qwen3: LoRA and QLoRA memory per model, and the license limits on derivatives.