# DeepSeek-V3 ![Organization](https://img.shields.io/badge/Org-DeepSeek-3b82f6) ![Context](https://img.shields.io/badge/Context-128K%20tokens-1f6feb) ![Pricing](https://img.shields.io/badge/Price-%240.27%20%2F%20%241.10%20per%201M-2da44e) ![License](https://img.shields.io/badge/License-MIT-2da44e) ![Weights](https://img.shields.io/badge/Weights-Open-2da44e) ![Released](https://img.shields.io/badge/Released-Dec%202024-555) > **Quick answer:** DeepSeek-V3 is DeepSeek's December 2024 flagship Mixture of Experts model, scoring 88.5% on MMLU, 90.2% on MATH, 27.2% on LiveCodeBench, and 34.5% on BigCodeBench. At $0.27/$1.10 per 1M tokens with MIT license and open weights, it was the leading cost-efficient frontier model at launch — matching GPT-4o-class performance at a fraction of the price. ## At a Glance **Where DeepSeek-V3 leads** - MIT license — fully open weights with no commercial restrictions - $0.27/$1.10 per 1M — among the lowest prices for GPT-4-class performance - 90.2% MATH — competitive with frontier models on mathematics - 88.5% MMLU — matches GPT-4 Turbo at 5× lower cost - 34.5% BigCodeBench — strong general coding capability - MoE architecture: 671B total, ~37B active — efficient inference **Where it lags** - 27.2% LiveCodeBench — below GPT-4.1, o-series models - 3.6% CyberGym — limited cybersecurity task performance - 24.9% SimpleQA — moderate factual accuracy - 128K context vs 1M for GPT-4.1 and Gemini 2.5 Pro - Superseded by DeepSeek-V3-0324 and later versions **Best for:** Cost-efficient frontier AI with open weights; self-hosting at GPT-4 capability level; high-volume pipelines where MIT licensing enables unrestricted deployment. ## What DeepSeek-V3 Is DeepSeek-V3 was released December 26, 2024, establishing DeepSeek as a leading provider of cost-efficient frontier AI. Its 671B MoE architecture (with ~37B active parameters at inference) delivers GPT-4-class performance at $0.27/$1.10 per 1M tokens — a breakthrough in cost/performance ratio that reshaped the competitive landscape. The model's MATH score (90.2%) and MMLU score (88.5%) placed it among the top tier of models at launch, competitive with GPT-4o and Claude 3.5 Sonnet at 10× lower price. Its MIT license allows unrestricted commercial use, self-hosting, and fine-tuning without attribution requirements. DeepSeek-V3 has since been updated (DeepSeek-V3-0324, then further versions), with each version improving benchmark scores while maintaining the same cost-efficient MoE architecture and open licensing approach. ## Specifications | Field | Value | |---|---| | Organization | DeepSeek | | Total parameters | 671B (MoE) | | Active parameters | ~37B per token | | Context window | 128,000 tokens | | License | MIT | | HuggingFace | [deepseek-ai/DeepSeek-V3](https://huggingface.co/deepseek-ai/DeepSeek-V3) | | Release date | December 26, 2024 | | Knowledge cutoff | October 2024 | | Modality | Text only | | Architecture | MoE (Mixture of Experts) | ## Pricing | | Input (per 1M tokens) | Output (per 1M tokens) | |---|---|---| | DeepSeek API | $0.27 | $1.10 | Open weights also available for self-hosting at no per-token cost. Pricing per [DeepSeek pricing page](https://platform.deepseek.com/usage). ## Context Window DeepSeek-V3 has a **128,000-token context window** — sufficient for most document analysis and coding tasks, though below the 1M-token context of GPT-4.1 and Gemini 2.5 Pro. ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [MMLU](/benchmarks/mmlu) | 88.5% | Benchgen evaluation | 2025-07 | | [MATH](/benchmarks/math) | 90.2% | Benchgen evaluation | 2025-07 | | [BigCodeBench](/benchmarks/bigcodebench) | 34.5% | Benchgen evaluation | 2025-07 | | [LiveCodeBench](/benchmarks/livecodebench) | 27.2% | Benchgen evaluation | 2025-07 | | [CyberGym](/benchmarks/cybergym) | 3.6% | Benchgen evaluation | 2025-07 | | [SimpleQA](/benchmarks/simpleqa) | 24.9% | Benchgen evaluation | 2025-07 | ## DeepSeek-V3 vs Alternatives | Model | MMLU | MATH | LiveCodeBench | Price (in/out per 1M) | License | |---|---|---|---|---|---| | **DeepSeek-V3** | 88.5% | 90.2% | 27.2% | $0.27 / $1.10 | MIT | | [GPT-4 Turbo](/models/openai/gpt-4-turbo) | 86.4% | 72.6% | 28.7% | $10 / $30 | Proprietary | | [GPT-4.1](/models/openai/gpt-4-1) | — | — | — | $2 / $8 | Proprietary | | [Llama 4 Maverick](/models/meta/llama-4-maverick) | — | — | — | Self-host | Llama 4 | DeepSeek-V3 vs GPT-4 Turbo: higher MMLU (88.5% vs 86.4%) and MATH (90.2% vs 72.6%) at 37× lower cost. The value case for DeepSeek-V3 is strong for any workload that fits within 128K context. ## Use DeepSeek-V3 via API ```python from openai import OpenAI client = OpenAI( api_key="YOUR_DEEPSEEK_API_KEY", base_url="https://api.deepseek.com/v1", ) response = client.chat.completions.create( model="deepseek-chat", messages=[{"role": "user", "content": "Solve this maths problem step by step..."}], ) print(response.choices[0].message.content) ``` ## Related Models & Benchmarks - Compare: [GPT-4.1](/models/openai/gpt-4-1) · [Llama 4 Maverick](/models/meta/llama-4-maverick) · [Qwen3 32B](/models/qwen/qwen3-32b) - Benchmarks: [MMLU](/benchmarks/mmlu) · [MATH](/benchmarks/math) · [BigCodeBench](/benchmarks/bigcodebench) · [LiveCodeBench](/benchmarks/livecodebench) ## Frequently Asked Questions

What is DeepSeek-V3?

DeepSeek-V3 is DeepSeek's December 2024 MoE flagship (671B total, ~37B active), scoring 88.5% MMLU and 90.2% MATH at $0.27/$1.10 per 1M tokens under MIT license with open weights.

Is DeepSeek-V3 open source?

Yes. DeepSeek-V3 is released under the MIT license on Hugging Face — free for commercial use, self-hosting, and modification with no restrictions.

What is DeepSeek-V3's context window?

DeepSeek-V3 supports a 128,000-token context window.

How does DeepSeek-V3 compare to GPT-4?

DeepSeek-V3 matches or exceeds GPT-4 Turbo on MMLU (88.5% vs 86.4%) and MATH (90.2% vs 72.6%) at $0.27/$1.10 vs $10/$30 per 1M tokens — approximately 37× lower cost for similar or better performance.

--- *Specs and scores sourced from DeepSeek's official V3 release (December 2024) and Benchgen evaluations. Pricing cited to the DeepSeek platform pricing page. Last updated 2026-07-24.*

DeepSeek-V3 — Evaluations

Evaluations