How to Deploy Qwen3 8B Locally
Qwen3 8B is a dense model with 8.2B parameters. At Q4 it needs at least 7GB of VRAM, and 12GB is the comfortable amount for an 8K context. The cheapest card that runs it today is the Tesla V100 16GB at about $659, at roughly 109.9 tokens/s.
Qwen3 8B details
How much VRAM does Qwen3 8B need for local deployment?
| Quantization | WeightsGB | Total at 8KGB | Total at 32KGB | Total at 128KGB |
|---|---|---|---|---|
| 1-bit | 2.4GGUF | 4 | 7.4 | — |
| 2-bit | 3.5GGUF | 5.2 | 8.5 | — |
| 3-bit | 4.3GGUF | 6 | 9.4 | — |
| 4-bit | 5GGUF | 6.7 | 10.1 | — |
| 5-bit | 5.9GGUF | 7.7 | 11 | — |
| 8-bit | 8.7GGUF | 10.6 | 14 | — |
| FP16 | 16.4GGUF | 18.1 | 21.5 | — |
Total = weights + KV cache (FP16) + 1GB runtime overhead. Weights marked GGUF are measured file sizes from the repository; the rest are estimated from bytes per parameter. Contexts beyond this model's 40K limit show a dash. Each tier takes whichever file of that bit width the repository has, preferring unsloth dynamic quants, so the exact file name varies by model.
Which GPUs can run Qwen3 8B?
AmazonJP
| Device | VRAMGB | Price | Est. t/s | Suggested context |
|---|---|---|---|---|
 Tesla V100 16GB | 16 | $6594 listings | 109.9 | 32K |
 Radeon RX 9070 XT | 16 | $8848 listings | 58.3 | 32K |
| GeForce RTX 5060 Ti 16GB | 16 | $9249 listings | 57.2 | 32K |
| GeForce RTX 5070 | 12 | $95610 listings | 83.9 | 32K |
| Radeon RX 7800 XT | 16 | $96610 listings | 56.9 | 32K |
| Tesla V100 32GB | 32 | $1,1413 listings | 109.9 | 32K |
 Arc Pro B60 | 24 | $1,1643 listings | 38 | 32K |
| Radeon RX 7900 XT | 20 | $1,2449 listings | 72.1 | 32K |
| Radeon RX 6950 XT | 16 | $1,4044 listings | 52.7 | 32K |
| GeForce RTX 5070 Ti | 16 | $1,57610 listings | 109.4 | 32K |
| Radeon RX 7900 XTX | 24 | $1,7389 listings | 85.5 | 32K |
| GeForce RTX 5080 | 16 | $1,8684 listings | 116.5 | 32K |
| Arc Pro B70 | 32 | $2,0743 listings | 50.2 | 32K |
| GeForce RTX 4080 | 16 | $2,1878 listings | 89.1 | 32K |
| Radeon AI PRO R9700 | 32 | $2,2446 listings | 58.3 | 32K |
| GeForce RTX 3090 | 24 | $2,5209 listings | 113.9 | 32K |
| Ryzen AI Max+ 395 128GB | 128 | $3,9136 listings | 24 | 32K |
| RTX PRO 4000 Blackwell | 24 | $3,9565 listings | 83.9 | 32K |
| GeForce RTX 4090 | 24 | $4,9059 listings | 121.8 | 32K |
| GeForce RTX 5090 | 32 | $7,4899 listings | 201.2 | 32K |
| DGX Spark 128GB | 128 | $8,2233 listings | 35.5 | 32K |
| GeForce RTX 2080 Ti 22GB | 22 | —No listings | 77.3 | 32K |
| GeForce RTX 5090 D V2 | 24 | —No listings | 157.2 | 32K |
| Arc Pro B65 | 32 | —No listings | 50.2 | 32K |
| GeForce RTX 4080 SUPER 32GB | 32 | —No listings | 91.3 | 32K |
| GeForce RTX 5090 D | 32 | —No listings | 201.2 | 32K |
| Instinct MI100 | 32 | —No listings | 107.4 | 32K |
| Instinct MI50 32GB | 32 | —No listings | 90.8 | 32K |
| Radeon PRO W7800 | 32 | —No listings | 52.7 | 32K |
| RTX PRO 4500 Blackwell | 32 | —No listings | 109.4 | 32K |
| A100 40GB PCIe | 40 | —No listings | 178.4 | 32K |
| CMP 170HX 40GB | 40 | —No listings | 178.9 | 32K |
| GeForce RTX 4090 48GB | 48 | —No listings | 121.8 | 32K |
| Radeon PRO W7900 | 48 | —No listings | 77.5 | 32K |
| RTX PRO 5000 Blackwell 48GB | 48 | —No listings | 157.2 | 32K |
| CMP 170HX 64GB | 64 | —No listings | 172.3 | 32K |
| Instinct MI210 | 64 | —No listings | 139.2 | 32K |
| RTX PRO 5000 Blackwell 72GB | 72 | —No listings | 157.2 | 32K |
| RTX PRO 6000D | 84 | —No listings | 179.7 | 32K |
| RTX PRO 6000 Blackwell | 96 | —No listings | 201.2 | 32K |
| Mac Studio M4 Max 128GB | 128 | —No listings | 54.9 | 32K |
| Mac Studio M5 Max 128GB | 128 | —No listings | 61.4 | 32K |
| MacBook Pro M4 Max 128GB | 128 | —No listings | 54.9 | 32K |
| MacBook Pro M5 Max 128GB | 128 | —No listings | 61.4 | 32K |
| Mac Studio M3 Ultra 512GB | 512 | —No listings | 80.6 | 32K |
Fits when weights + KV cache + runtime overhead is at or below usable memory, single card. "Offload" on MoE models means it runs with expert weights in system RAM (only attention layers and KV cache stay in VRAM); that speed assumes 70 GB/s RAM bandwidth. The context column is the largest tier that fits.
Deploy Qwen3 8B locally with Ollama, llama.cpp or vLLM
Ollama pulls and runs it in one command; the quantization comes from the official tag.
ollama run qwen3:8b
llama.cpp pulls the GGUF straight from Hugging Face; -c sets the context length.
llama-server -hf Qwen/Qwen3-8B-GGUF:Q4_K_M -c 8192
vLLM serves the original-precision weights, which needs far more memory than GGUF and usually more than one GPU.
vllm serve Qwen/Qwen3-8B
File names follow that week's model-library snapshot; the definitions are in the methodology. Methodology
FAQ
How much VRAM does Qwen3 8B need?
At Q4_K_M with an 8K context Qwen3 8B needs about 7GB of VRAM; 12GB leaves comfortable headroom.
What is the cheapest GPU that runs Qwen3 8B?
The Tesla V100 16GB: 16GB of VRAM, currently about $659, at roughly 109.9 tokens/s.
Can you run Qwen3 8B with Ollama?
Yes: ollama run qwen3:8b.