Deepseek
Open-source reasoning LLM
Frontier-grade reasoning at open-source, pennies-per-token cost
Reviewed by The Desk · Last verified July 2026
“DeepSeek proved a frontier-adjacent reasoning model could ship with open MIT weights at a fraction of incumbents' prices, and that's still its core appeal: coding and STEM performance close to the closed leaders for pennies per million tokens, or free in the app. The catch is trust, not capability — the hosted service runs on Chinese infrastructure, censors sensitive topics, and documents little about privacy, so serious enterprise use realistically means self-hosting the weights.”
Skip it ifYou handle regulated or sensitive data and can't send prompts to China-based servers, or you need enterprise support, SLAs, and compliance guarantees — in which case self-hosting the open weights is the only safe path, and that takes real GPU budget and MLOps effort.
What it is
DeepSeek is a family of large language models from the Chinese lab of the same name, backed by quant hedge fund High-Flyer and led by founder Liang Wenfeng. Its headline products are reasoning-focused models that expose their chain of thought, delivered three ways: a free consumer chatbot on web and mobile, a low-cost developer API, and downloadable open weights you can run yourself.
The pitch that made it famous in early 2025 is cost. DeepSeek trained models that trade blows with closed frontier systems on coding and math benchmarks, priced the API at a fraction of competitors, and released the flagship weights under a permissive MIT license. For developers that means prototyping against the hosted API for pennies, then optionally self-hosting the same model for full data control - a combination the incumbents don't offer.
The trade-offs are about trust rather than capability. The hosted app and API run on servers in China, censor politically sensitive subjects, and document little about how prompts are retained, which rules the managed service out for many regulated or enterprise buyers. "Open" also has limits: the weights are public but the training data and pipeline are not, so the models can't be fully audited or reproduced. For cost-sensitive builders and self-hosters DeepSeek is genuinely compelling; for compliance-bound teams, the only comfortable path is running the weights on their own infrastructure.
Key features
- Reasoning models that expose their chain of thought on math, logic, and code
- Open MIT-licensed weights you can download, fine-tune, and self-host
- API pricing an order of magnitude below closed frontier models
- Free, unlimited chat on web and iOS/Android with no subscription
- Automatic context caching that cuts repeat-prompt costs
- OpenAI-compatible API that drops into existing tooling
Who it’s for
Writing and debugging code with chain-of-thought reasoning
Solving math, logic, and multi-step STEM problems
Powering apps with a low-cost API backend
Self-hosting a private LLM on your own GPUs for data control
Pros & cons
The good
- Dramatically cheaper than comparable frontier models, both on the API and as a free app
- Genuinely open MIT-licensed weights enable private self-hosting and fine-tuning
- Reasoning and coding quality that trades blows with far pricier closed models
The catch
- Hosted app and API run on China-based servers with little documented privacy protection, a nonstarter for many regulated buyers
- Censors politically sensitive topics on the managed service
- Training data and full pipeline are not released, so 'open' has real limits, and enterprise tooling/support lag OpenAI and Anthropic
Pricing
| Tier | Price | What you get |
|---|---|---|
| Web & Mobile Chat | Free | Unlimited chat · Reasoning mode · iOS and Android apps |
| API - Flash tier | Custom | Usage-based per token · ~$0.14 per 1M input tokens (cache miss) · ~$0.28 per 1M output tokens · OpenAI-compatible endpoint |
| API - Pro tier | Custom | Higher-capability model · Usage-based per token |
| Self-hosted (open weights) | Free | MIT-licensed weights · Runs on your own GPUs · Full data control · Fine-tunable |
Free tier — Unlimited free web and mobile chat; MIT-licensed weights free to self-host
Alternatives to Deepseek
More Chatbots & LLMs →
Questions people ask
Is DeepSeek really open source?
The weights for its flagship reasoning models are released under the permissive MIT license, so you can download, fine-tune, self-host, and use them commercially. However, DeepSeek has not released the training data or full training code, so it's more accurately open-weight than fully reproducible open source.
How much does DeepSeek cost?
The web and mobile chat apps are free. The API is usage-based and among the cheapest available - roughly $0.14 per million input tokens (cache miss) and $0.28 per million output on the Flash tier, with a higher-capability Pro tier costing more. Model names and prices shift often, so check the official docs before committing.
Is DeepSeek safe for business or sensitive data?
The hosted app and API run on China-based servers with limited documented privacy protections, and the managed service censors politically sensitive topics, which is a dealbreaker for many regulated or enterprise users. Self-hosting the open weights keeps data entirely on your own infrastructure and sidesteps those concerns.
How does DeepSeek compare to ChatGPT or Claude?
On coding and STEM reasoning benchmarks it's competitive with the closed frontier leaders at a fraction of the price, but it lags on enterprise tooling, ecosystem breadth, support, and safety guardrails.