DeepSeek V3.
DeepSeek's flagship MoE — 671B total, 37B active, frontier-class.
2× AMD MI325.
Most-aggressive quantisation we have a working recommendation for. Lower precision = less VRAM = cheaper hardware, at a small accuracy cost.
Cheapest hosted endpoints.
What it's best at.
Speed across providers.
Tokens/sec and time-to-first-token measured against the same prompt template on each provider's API.
| Provider | Tokens/sec | TTFT | Total |
|---|---|---|---|
| OpenRouter | 30.3 | 1207 ms | 4324 ms |
Variants in the DeepSeek family.
DeepSeek R1
DeepSeek's reasoning model — RL-trained, frontier-class, MIT-licensed.
DeepSeek R1 Distill Llama 70B
70B Llama distilled from DeepSeek R1's reasoning traces.
DeepSeek R1 Distill Qwen 32B
32B Qwen base distilled from DeepSeek R1.
DeepSeek R1 Distill Qwen 14B
14B distilled R1 — laptop-friendly reasoning.
DeepSeek R1 Distill Qwen 7B
7B distilled R1 — runs on any modern GPU.
DeepSeek R1 Distill Qwen 1.5B
Tiny distilled R1 — phone / browser deployable.
Frequently asked.
How do I run DeepSeek V3?
Where can I access DeepSeek V3?
How much does it cost to run DeepSeek V3?
Is DeepSeek V3 open-source or proprietary?
Cheapest hardware per quantisation.
Each row is one quantisation tier (the same weights compressed differently). Lower precision → lower VRAM → cheaper hardware, at the cost of small accuracy loss. $/hr refreshed hourly from each provider's API.
| Quantisation | Cheapest GPU config | Total VRAM | Live $/hr | tokens/sec | |
|---|---|---|---|---|---|
|
FP16
FP16 — half precision (default)
|
2048 GB | — | — | Compare → | |
|
FP8
FP8 — 8-bit float (Hopper / Blackwell)
|
1024 GB | — | — | Compare → | |
|
INT8
INT8 — 8-bit integer
|
640 GB | $3.84/hr | — | Compare → | |
|
INT4
INT4 — 4-bit integer (~4× VRAM saving)
|
512 GB | — | — | Compare → |
What it costs per month across providers.
Estimate your monthly bill for DeepSeek V3 across every host that publishes per-token pricing. Slide your token volumes; the chart + table re-rank cheapest-first.
Cheapest provider on the left.
Total monthly cost — input + output tokens combined.
Bill breakdown.
Rent the GPU instead of paying per token.
For an open-weights model like DeepSeek V3, you can rent a GPU and serve inference yourself. The math: cheapest GPU rental × 730 hours/month + your electricity rate × power draw.
Assumes the GPU runs 24/7 at ~85% utilisation. If your traffic is bursty, you'll pay less for the API and probably more for the GPU (idle hours still cost rental). The breakeven analysis lives on the Self-host vs API breakeven tool.
What it's best at.
Scores normalised against benchmark ceilings (100 = perfect). Coloured by tier — coral 80+ frontier, lavender 65+ strong, sage 50+ solid, slate below.
Published scores.
| Benchmark | Score | Source |
|---|---|---|
| GPQA | 59.1 | official ↗ |
| MATH | 90.2 | official ↗ |
| MMLU | 88.5 | official ↗ |
| MMLU-Pro | 75.9 | official ↗ |
| HumanEval | 82.6 | official ↗ |
Independent rankings.
About DeepSeek V3.
DeepSeek V3 is DeepSeek AI's flagship MoE model — 671B parameters total, 37B activated per token, trained on 14.8T tokens for an estimated $5.5M (vastly cheaper than comparable frontier models, made possible by FP8 mixed-precision training and Multi-head Latent Attention). Benchmark performance rivals GPT-4o and Claude 3.5 Sonnet at a fraction of the inference cost. Available on Hugging Face under DeepSeek's permissive commercial license. The V3 release in late 2024 was a watershed moment for open-weight models — it proved frontier capability didn't require frontier budgets.
How it's built.
How much it can remember.
What it can do.
Every place this model is hosted.
Self-hosted on rented GPU cluster
self hostedDeepSeek API
api directAmong cheapest frontier-class APIs