deepseek-v3

DeepSeek V3 is DeepSeek's flagship general-purpose model — a 671B mixture-of-experts model (37B active parameters per token) trained with a novel multi-token prediction objective and auxiliary-loss-free load balancing. It is not a reasoning model (unlike DeepSeek-R1) but a fast, capable general-purpose model competitive with frontier proprietary models on many tasks. It is open-weights (MIT license) and available via DeepSeek's API and multiple third-party providers.

DeepSeek V3 is priced at $0.27 per million input tokens (cache miss) and $1.10 per million output tokens from DeepSeek's API, with cache hits at $0.07/1M input — among the lowest prices for a frontier-class model. At 1,000 input + 500 output tokens per call, it costs approximately $0.0008 per 1,000 API calls — among the lowest prices for a frontier-class general-purpose model. Third-party providers offer similar rates. Self-hosted deployments require 8× H100 80GB GPUs for the full model, but the MoE architecture makes inference more efficient than a dense model of equivalent benchmark performance.

DeepSeek V3 supports a 128K-token context window. API latency from DeepSeek's own endpoints is approximately 300–800ms TTFT under normal load, though reliability has been inconsistent. Third-party providers offer more stable latency. Throughput is approximately 40–80 tokens per second. Maximum output is 8,192 tokens. The MoE architecture (37B active parameters) means inference cost and speed are closer to a 37B dense model despite the 671B total parameter count — enabling competitive throughput on standard GPU hardware.

DeepSeek V3 is best suited to general-purpose language tasks where cost efficiency is important: coding, document analysis, summarization, Q&A, and content generation. On coding benchmarks, it performs comparably to leading proprietary models at a fraction of the cost — making it particularly attractive for code generation pipelines. Its MIT license enables commercial use, fine-tuning, and self-hosting without restrictions. For teams with high-volume coding or general NLP workloads, V3 offers an exceptional cost-to-quality ratio. It is not a reasoning model — for tasks requiring extended chain-of-thought, DeepSeek-R1 is more appropriate.

DeepSeek V3 sits in the advanced tier of DeepSeek's lineup, below DeepSeek-R1 in reasoning capability but faster and cheaper for general tasks. Its main competitors are proprietary frontier models from OpenAI and Anthropic (more expensive, no self-hosting), and open-weights models with multimodal capability. The primary concerns for production use are API reliability (DeepSeek's infrastructure has had outages) and data privacy (Chinese data regulations apply to DeepSeek's API). For teams using US-based third-party providers or self-hosting, these concerns are mitigated. Its intelligence score of approximately 84/100 reflects strong advanced-tier performance at an exceptional price point.