Alibaba
Qwen 3 Coder-Next (80B)
Efficient MoE coding model for local and hosted coding agents
Model Summary
Family
Qwen
Version
3.0
Parameters
80B (est.)
Parameter counts for closed models are estimates; vendors rarely publish exact sizes.
VRAM Requirements by Quantization
Memory needed to serve Qwen 3 Coder-Next (80B) for inference, including a 20% overhead for activations and KV cache.
| Precision | VRAM needed | Smallest single GPU that fits |
|---|---|---|
| INT4 (4-bit) | 44.70 GB | NVIDIA L40S (48 GB) |
| INT8 (8-bit) | 89.41 GB | NVIDIA RTX PRO 6000 Blackwell (96 GB) |
| FP16 (16-bit) | 178.81 GB | NVIDIA B200 SXM (192 GB) |
| FP32 (32-bit) | 357.63 GB | Multi-GPU required |
Recommended GPU Configurations
Cheapest on-demand configurations to serve Qwen 3 Coder-Next (80B) at 8-bit (89 GB VRAM).
1x AMD Instinct MI250X
128 GB total VRAM · CDNA 2
~$2.50/h
6x NVIDIA T4
96 GB total VRAM · Turing
~$3.00/h
6x NVIDIA P100 SXM2
96 GB total VRAM · Pascal
~$3.60/h
Quick GPU Planning
Use the calculator pre-filled with this exact version to estimate memory, speed, and compute requirements in a few clicks.
Access Pre-filled CalculatorFrequently Asked Questions
How much VRAM do you need to run Qwen 3 Coder-Next (80B)?
With an estimated 80B parameters, Qwen 3 Coder-Next (80B) needs roughly 89 GB of VRAM in 8-bit (INT8), 45 GB in 4-bit, and 179 GB in FP16, including a 20% overhead for activations and KV cache.
Which GPUs can run Qwen 3 Coder-Next (80B)?
At 8-bit quantization, the most cost-effective option is 1x AMD Instinct MI250X (128 GB combined VRAM, around $2.50/hour on-demand). Higher-end cards like the NVIDIA B200 or AMD MI355X reduce the GPU count needed.
Can Qwen 3 Coder-Next (80B) run on a single GPU?
Yes. In 8-bit, a single NVIDIA RTX PRO 6000 Blackwell (96 GB) fits the model.
How much does it cost to serve Qwen 3 Coder-Next (80B) in the cloud?
Renting 1x Instinct MI250X costs on the order of $2.50/hour, i.e. about $1,825/month running 24/7. Actual prices vary by provider and commitment; spot and reserved capacity can be significantly cheaper.
Deploy on a GPU cloud
Rent 1x Instinct MI250X by the hour instead of buying hardware.
