DeepSeek-V3
DeepSeek-V3 is an open-weights language model from DeepSeek, released December 26, 2024. It has 684B parameters (mixture-of-experts, 51B active). The context window is 160K tokens. On the Epoch Capabilities Index it scores 131.7, rank 154 of 256 rated models.
Check it against your own PC →
Can you run DeepSeek-V3 locally?
Even with 32 GB of VRAM and 32 GB of RAM it is too large to run at a usable speed on a single consumer GPU. Estimates assume 32 GB of DDR5 RAM, a 16K context and llama.cpp-style GGUF quantization. Enter your exact hardware in the live app.
| GPU memory | Verdict | Best quant | VRAM | RAM | tok/s |
|---|---|---|---|---|---|
| 6 GB VRAM | Too large | – | – | – | – |
| 8 GB VRAM | Too large | – | – | – | – |
| 12 GB VRAM | Too large | – | – | – | – |
| 16 GB VRAM | Too large | – | – | – | – |
| 24 GB VRAM | Too large | – | – | – | – |
| 32 GB VRAM | Too large | – | – | – | – |
Ready-made quantizations on Hugging Face: unsloth/DeepSeek-V3-GGUF · bullerwins/DeepSeek-V3-GGUF · mmnga/DeepSeek-V3-bf16-gguf · QuixiAI/DeepSeek-V3-AWQ · v2ray/DeepSeek-V3-1B-Test-AWQ · OPEA/DeepSeek-V3-int4-sym-awq-inc
Strengths by category
ModelTape's 0–100 fit on Epoch AI benchmark results.
| Category | Score | Based on |
|---|---|---|
| Math | 16.3 | MATH level 5, OTIS Mock AIME 2024-2025, FrontierMath-2025-02-28-Private |
| Knowledge | 41.0 | GPQA diamond, MMLU |
Benchmark results
| Benchmark (Epoch AI) | Score |
|---|---|
| MMLU | 82.9% |
| MATH level 5 | 64.9% |
| Aider Polyglot | 48.4% |
| GPQA Diamond | 43.8% |
| OTIS Mock AIME | 15.7% |
| FrontierMath-2025-02-28-Private | 3.0% |
| SimpleBench | 2.7% |
Links
Hugging Face: deepseek-ai/DeepSeek-V3 · Epoch AI benchmarks
Models with a similar score
Gemini 2.0 Pro · o1-mini · DeepSeek-V3 (Mar 2025) · Qwen3-30B-A3B-Instruct (Jul 2025) · gpt-oss-20b · GPT-4.1 mini · Qwen3-30B-A3B · Gemini 2.0 Flash (Feb 2025)
More from DeepSeek
DeepSeek V4.1 Flash 1017 · DeepSeek V4 Flash Vision Exp · DeepSeek V4.1 Flash · DeepSeek V4 Pro 0813 · DeepSeek V4 Flash 0731 · DeepSeek V4 Pro DSpark · DeepSeek V4 Flash DSpark · DeepSeek-V4-Pro