DeepSeek-R1 was released January 20, 2025 with a technical report, downloadable weights and an API route. The provider’s launch identified the API selection at that time as `deepseek-reasoner`; this is a hosted routing name rather than a guarantee that all calls using it in later years run exactly the January 2025 checkpoint.
Historical research cutoff: 25 October 2026. Primary-source historical claims are date-scoped; cloud entitlements, region-specific availability, current pricing and rate limits are not implied.
Release identity at a glance
- Provider
- DeepSeek
- Model family
- DeepSeek-R1
- Public launch
- 2025-01-20
- API ID / checkpoint
- deepseek-reasoner
- API retirement status
- Not confirmed as retired at the research cutoff
The release and why it mattered
The release distinguished the full R1 model from distilled smaller descendants. DeepSeek’s repository describes a 671-billion-parameter mixture-of-experts model with 37 billion activated parameters and a 128K context in its reference configuration. Distilled Qwen- and Llama-based variants are separate artifacts with license histories influenced by their base models.
Capabilities and genuine use cases
R1 demonstrated how inference-time reasoning methods could be combined with publicly distributed weights. For users, the choice is between operating a weight checkpoint locally, paying a hosted API, or deploying a distilled model. Each path changes memory needs, update control, data governance and testing burden; an OpenAI-compatible HTTP shape does not mean equivalent response semantics.
Integration and version identity
For local deployment, use the repository’s exact model files and verify the hashes, quantization, serving engine and context settings. The reference README discusses local deployment options and derivative model details. For hosted use, inspect DeepSeek’s dated changelog because a mutable `deepseek-reasoner` alias can change model backing while preserving the endpoint name.
Limits and evidence gaps
The repository states MIT licensing for the R1 series but also explains separate inheritance for some distilled variants. Do not generalize “MIT” to every related base model without reading each artifact’s license. DeepSeek’s headline benchmark comparisons are vendor claims and require independent reproduction before comparative rating.
How to choose or migrate
DeepSeek-R1 is important both for reasoning research and open-weight deployment. Buyers should balance measured answer quality with infrastructure, security review and license terms. Do not copy the January 2025 token rates into an October 2026 cost comparison; AI 10 owns that record.
Reproducible evaluation checklist
When using weights, record repository revision, artifact hash, serving engine, quantization and base-model license. With a hosted API, also log the exact serving alias and the dated changelog entry that describes its backing checkpoint.
Perform multilingual, extraction and security validation with the same input set across local and hosted deployments. Similar HTTP endpoints do not imply the same tool protocol or refusal behavior.
Sources & Further Reading
- deepseek.com — documentation/reference (supplier-provided reference; not independently checked in this pass)
- github.com — documentation/reference (supplier-provided reference; not independently checked in this pass)
- DeepSeek API changelog (retrieved from first-party public documentation on 2026-10-25; this is not a deployment test)
Source documentation establishes vendor-published facts, not independently measured model quality, tenant access, contractual rights or current prices. Use a dated pricing ledger and an account-specific entitlement check for deployment decisions.