DeepSeek V3
by DeepSeek
DeepSeek V3 is a 685B Mixture-of-Experts model with only 37B active parameters per forward pass, delivering frontier-class performance on coding and general tasks. It offers the best price-to-performance ratio among non-reasoning models at $0.27/M input tokens.
Send a message to start the conversation.
More from DeepSeek
View all →DeepSeek R1
DeepSeek R1 is a 671B open-source reasoning model trained with reinforcement learning, matching or beating OpenAI o1 on math, science, and coding benchmarks. At $0.55/M input tokens versus o1's $15, it delivers frontier reasoning at a 96% cost reduction.
$0.55 / M tokens
DeepSeek R1 Zero
DeepSeek R1 Zero is the base RL-trained checkpoint of DeepSeek R1 before supervised fine-tuning, demonstrating raw chain-of-thought reasoning emerging directly from reinforcement learning. It's primarily used for research into reasoning emergence and RL-based training methods.
$0.40 / M tokens
DeepSeek Coder V2
DeepSeek Coder V2 is a 236B MoE coding model with 21B active parameters, achieving state-of-the-art performance on HumanEval and LiveCodeBench among open models. At $0.14/M input tokens, it offers frontier code generation at a price comparable to much smaller models.
$0.14 / M tokens
Janus Pro
Janus Pro is DeepSeek's unified multimodal model that uses a dual-encoder architecture to separately handle visual understanding and image generation. It delivers strong prompt alignment and competitive quality relative to its low price.
$0.010 / image