The lab that proved compute isn't everything.
The Lab
Hangzhou-based, founded in 2023, DeepSeek shocked Silicon Valley in early 2025 by training a frontier-class reasoning model for what was reported as a tiny fraction of OpenAI's cost. The technical innovation was the mixture-of-experts architecture done right: 671B total parameters but only 37B active per token, with engineering tricks that compressed training compute dramatically.
The 2026 strategy is to ride that efficiency. DeepSeek R2 is the open-weights reasoning model that matches GPT-5.4 Think on math and competitive programming at roughly 1/30th the price. V3.5 is the general-purpose flagship at near-Haiku pricing with near-Sonnet quality. Both are downloadable, both are permissively licensed, both have already been forked dozens of times by the open-source community.
The catch mirrors Qwen's. Western enterprise customers face the same procurement caution about Chinese-origin AI. DeepSeek's response has been to ship faster and cheaper than rivals can ignore — for the open-source community, for the cost-sensitive market, and for everyone who watched the early-2025 cost story and concluded that compute spending isn't quite the moat the US labs claimed.
Recent Coverage
Models
Open-weights reasoning model that matches GPT-5.4 Think on math and competitive programming at roughly 1/30th the price. The model researchers benchmark against.
Best for
General-purpose flagship. Sparse MoE with 671B total params, 37B active per token. Quality close to Sonnet-class, cost close to Haiku-class. The price-quality outlier of 2026.
Best for
Code-tuned variant of V3. The default open-weights pick for self-hosted coding agents in the open-source community.
Best for
When to Pick DeepSeek · When to Pick Someone Else
Each lab in the atlas comes with its own positioning, model line, and use cases. The point of organising the AI landscape by lab is that the answer to "which model should I use" almost always starts with "which lab is closest to what I'm trying to do."
