Table of Contents
DeepSeek-V2 16B Lite is a Mixture-of-Experts model with 16B total params and 2.4B active per token. The MoE architecture means it runs faster than a dense 16B model would suggest — but it occupies the full VRAM of the larger weight footprint. This page lists exact cost-per-token numbers across every GPU we rent.
RTX 5090 + FP8 at £0.24/1M tokens is the cost leader for DeepSeek-V2 16B Lite. RTX 4090 at AWQ-INT4 is the cheapest path on Ada hardware. For DeepSeek-V2 236B, only multi-GPU configs work; expect £1.50–2.50/1M tokens self-hosted.
Methodology
Standard methodology — 60% utilisation, vLLM 0.6.3 with continuous batching, 50-thread Locust driver, Ubuntu 22.04. Models from deepseek-ai/DeepSeek-V2-Lite-Chat (BF16) and the AWQ-INT4 community port.
DeepSeek-V2 16B Lite by GPU
| GPU | Monthly | Precision | tok/s | Cost per 1M |
|---|---|---|---|---|
| RTX 5060 Ti 16 GB | £119 | AWQ-INT4 | 320 | £0.34 |
| RTX 5080 | £189 | AWQ-INT4 | 520 | £0.28 |
| RTX 3090 | £159 | AWQ-INT4 | 410 | £0.28 |
| RTX 4090 | £289 | AWQ-INT4 | 590 | £0.30 |
| RTX 5090 | £399 | AWQ-INT4 | 780 | £0.30 |
| RTX 5090 | £399 | FP8 | 950 | £0.24 |
| RTX 6000 Pro | £899 | FP16 | 1,420 | £0.50 |
RTX 5090 with FP8 is the cost leader. RTX 3090 is the cheapest dedicated server with sufficient VRAM for AWQ-INT4.
DeepSeek-V2 236B (multi-GPU only)
V2 236B has 21B active params per token but the full 236B has to be VRAM-resident. ~470 GB FP16 weights makes it multi-GPU territory exclusively.
| Config | Monthly | Precision | tok/s | Cost per 1M |
|---|---|---|---|---|
| 4× RTX 5090 | £1,800 | AWQ-INT4 (TP=4) | 180 | £1.93 |
| 2× RTX 6000 Pro | £2,200 | FP8 (TP=2) | 210 | £2.02 |
| 4× A100 80 GB | POA | FP16 (TP=4) | 270 | POA |
| 8× H100 | POA (£15K+) | FP8 (TP=8) | 480 | ~£3.50 |
For V2 236B the official DeepSeek API at £0.18/1M is hard to beat unless you need data residency. We size custom builds on request.
DeepSeek-Coder variants
| Variant | Params | VRAM (INT4) | tok/s on 5090 | Cost per 1M |
|---|---|---|---|---|
| DeepSeek-Coder-V2-Lite | 16B (2.4B active) | 10 GB | 780 | £0.30 |
| DeepSeek-Coder 6.7B | 6.7B | 4 GB | ~1,800 | £0.13 |
| DeepSeek-Coder 33B | 33B | 18 GB | ~310 | £0.74 |
| DeepSeek-Coder-V2 236B | 236B | Multi-GPU | cluster | £1.50+ |
Cost leaders
- V2 16B Lite cost leader: RTX 5090 + FP8 at £0.27/1M.
- V2 16B Lite budget option: RTX 3090 / 4090 at AWQ-INT4 — slightly higher cost-per-token but cheaper monthly.
- DeepSeek-Coder 6.7B: RTX 5090 at FP16 — £0.14/1M tokens, fastest cost leader of the lot.
- V2 236B: use the official DeepSeek API unless you have a residency requirement.
Bottom line
For typical DeepSeek deployments, the RTX 5090 at FP8 is the right host. Drop to a 4090 or 3090 if FP8 is not available. For V2 236B and V3 671B, hosted APIs win on cost — see cost: run DeepSeek vs API for the break-even math.
