Open-source LLM · DeepSeek
DeepSeek V3
671 B parameters (37 B active per token) · 128K context · MIT · released 2024-12-26 by DeepSeek.
DeepSeek V3 needs about 1342 GB at FP16: no single GPU can hold it. Serving requires an 8-GPU HGX baseboard (INT8/INT4) or a full rack (NVL72, TPU pod) at FP16.
Where this model sits
Frontier open-weight of late 2024. MoE means only 37B parameters activate per token, so inference cost is 8B-class despite the 671B footprint on disk.
What runs it (production serving)
Datacenter GPUs and rack designs sized for DeepSeek V3. Consumer cards are not practical at this size.
Smallest datacenter GPU that fits (INT4)
Multi-GPU / rack-scale (FP16)
Memory at different context lengths
Weights are fixed. The KV cache grows linearly with context length and can rival the weights at full window. Most guides only quote weights, then hand-wave a 15% overhead, which is only true at short context.
| Context | FP16 total | INT8 total | INT4 total |
|---|---|---|---|
| 4K (short chat) | 1342 + 15.3 = 1357.3 GB | 671 + 7.6 = 678.6 GB | 336 + 7.6 = 343.1 GB |
| 32K (typical RAG) | 1342 + 122.0 = 1464.0 GB | 671 + 61.0 = 732.0 GB | 336 + 61.0 = 396.5 GB |
| 128K (full window) | 1342 + 488.0 = 1830.0 GB | 671 + 244.0 = 915.0 GB | 336 + 244.0 = 579.5 GB |
KV bytes per token = 2 × 61 layers × 128 KV heads × 128 head-dim × precision bytes. INT4 rows keep the KV cache at INT8 (typical serving default).
Quality benchmarks
Vendor-reported scores from the release blog or tech report. Every row cites its source.
| Benchmark | Score | Source | Date |
|---|---|---|---|
| MMLU (5-shot) | 88.5 | DeepSeek V3 tech report | 2024-12-26 |
| HumanEval | 82.6 | DeepSeek V3 tech report | 2024-12-26 |
| MATH (0-shot) | 61.6 | DeepSeek V3 tech report | 2024-12-26 |
| GPQA | 59.1 | DeepSeek V3 tech report | 2024-12-26 |
Speed and cost to serve
What cloud providers currently charge to serve DeepSeek V3. Tokens per second is what the provider publishes on their own page. Batch size, prompt length, and inference engine can shift these numbers a lot; treat as directional.
| Provider | Hardware | Speed | $/M in | $/M out | As of |
|---|---|---|---|---|---|
| DeepSeek API | H800 | — | $0.27 | $1.10 | 2025-08-01 |
| Together AI | H100 nodes | — | $1.25 | $1.25 | 2025-08-01 |
| Fireworks AI | H100 nodes | — | $0.90 | $0.90 | 2025-08-01 |
Cheapest option in this list for output-heavy workloads: Fireworks AI at $0.90/M out. Fastest published: not enough data.
License and variants
- License
- MIT
- The gotcha
- MIT weights, but the DeepSeek terms of use restrict certain military and surveillance applications.
- Variants
- Base · Chat
- Context window
- 131,072 tokens (128K)
- Architecture
- MoE (671B total, 37B active per token)
Compare with
Model card: https://huggingface.co/deepseek-ai/DeepSeek-V3 · Release: DeepSeek announcement
See also: every open-source LLM we track · all AI chips · rack designs.