DeepSeek

Yes — the web app and mobile app are free. Paid usage only kicks in when you call the API or self-host on your own GPUs.

DeepSeek trained a Mixture-of-Experts (MoE) model that activates only ~37B of its 671B parameters per token, so inference cost is a fraction of dense frontier models. Public weights let cloud vendors compete on price.

V3 is the general-purpose chat/coding model. R1 is the reasoning model with visible chain-of-thought; use R1 for math, code review, planning and hard reasoning, use V3 for everything else.

DeepSeek's Chinese writing and reasoning are noticeably more natural than ChatGPT on native mainland-Chinese prompts. For English long-form writing, Claude still leads.

Yes. Full weights are released under an MIT-style license. Practical self-hosting requires 8×H100 or equivalent for the full V3/R1; distilled variants run on a single 24 GB GPU.

The hosted API stores prompts and may use them for service improvement unless you opt out. For sensitive data, deploy the open-source weights on your own infrastructure or use a private cloud (Azure, AWS, Volcano Engine) that offers DeepSeek without data retention.

V3 is competitive with Claude 3.5 Sonnet on HumanEval and beats GPT-4o on many benchmarks. R1 excels at multi-file refactor reasoning. Cursor and Cline both support DeepSeek as the underlying model.

V3 has limited multimodal support; DeepSeek-VL handles vision separately. For serious multimodal work, Gemini 2.5 Pro or GPT-5 remain stronger choices in 2026.

The hosted DeepSeek service filters politically sensitive Chinese topics. The open-source weights on your own hardware do not carry the same filter but still reflect training-data bias.

Teams needing tightly integrated voice, image and plugin ecosystems (ChatGPT Plus wins) or the most natural English long-form writing (Claude wins). For everyone else it is the best price/performance model in 2026.