DeepSeek-R1

Checkpoint identity

Official identifier

deepseek-ai/DeepSeek-R1DeepSeek-R1 official model card (external source) · DeepSeek-R1 checkpoint configuration (external source) · DeepSeek-R1 MIT license (external source) · DeepSeek-V3 architecture and weight notes (external source)

Architecture

Mixture-of-experts reasoning language model based on DeepSeek-V3-BaseDeepSeek-R1 official model card (external source) · DeepSeek-R1 checkpoint configuration (external source) · DeepSeek-R1 MIT license (external source) · DeepSeek-V3 architecture and weight notes (external source)

Total parameters

671 billionPublished rounded counts; total stored weights determine the preliminary footprint, including all MoE experts.DeepSeek-R1 official model card (external source) · DeepSeek-R1 checkpoint configuration (external source) · DeepSeek-R1 MIT license (external source) · DeepSeek-V3 architecture and weight notes (external source)

Active parameters

37 billion per tokenActive parameters are a compute-related count, not the model's full stored weight count.DeepSeek-R1 official model card (external source) · DeepSeek-R1 checkpoint configuration (external source) · DeepSeek-R1 MIT license (external source) · DeepSeek-V3 architecture and weight notes (external source)

Modalities

text-input, text-outputDeepSeek-R1 official model card (external source) · DeepSeek-R1 checkpoint configuration (external source) · DeepSeek-R1 MIT license (external source) · DeepSeek-V3 architecture and weight notes (external source)

Context

Not verified for this recordNative configuration where directly verified. Extensions, advertised K-labels and configuration differences are preserved in the notes below.DeepSeek-R1 official model card (external source) · DeepSeek-R1 checkpoint configuration (external source) · DeepSeek-R1 MIT license (external source) · DeepSeek-V3 architecture and weight notes (external source)

Release date

Exact release date not independently established in the reviewed artifactsDeepSeek-R1 official model card (external source) · DeepSeek-R1 checkpoint configuration (external source) · DeepSeek-R1 MIT license (external source) · DeepSeek-V3 architecture and weight notes (external source)

Download access

Public checkpoint referenceDeepSeek-R1 official model card (external source) · DeepSeek-R1 checkpoint configuration (external source) · DeepSeek-R1 MIT license (external source) · DeepSeek-V3 architecture and weight notes (external source)

License

MITThe original full DeepSeek-R1 checkpoint and code are declared MIT. Preserve the copyright and permission notices. Distilled Qwen and Llama checkpoints have additional source-model license considerations and are not this record.DeepSeek-R1 official model card (external source) · DeepSeek-R1 checkpoint configuration (external source) · DeepSeek-R1 MIT license (external source) · DeepSeek-V3 architecture and weight notes (external source)

Theoretical weight-memory footprint

Using the publisher's approximately 671 billion total parameters at 16 bits per parameter gives 1,342 GB (1249.8 GiB) of nominal weights.

This estimate excludes KV cache, activations, quantization metadata, non-weight buffers and serving overhead. It is not a tested GPU fit. MoE routing activates only part of the network per token while the complete expert weights still need storage.

Quantization

The checkpoint configuration specifies blockwise FP8. Conversion or lower precision needs an explicitly supported backend and an independent quality evaluation.

Serving considerations

• The card documents approximately 671B main-model parameters and 37B activated per token. Additional multi-token prediction weights make the downloaded artifact count larger; active count is not resident memory size.
• The card advertises 128K context tokens, while config.json has max_position_embeddings 163,840. The advertised context is preserved here without treating that config limit as a tested usable context; no normalized context integer is asserted.
• This is the original R1 checkpoint, not R1-0528 or a distilled model. Reasoning output can lengthen generation. Aguasto has not evaluated or deployed it.

Framework documentation

These are third-party documentation starting points. Support depends on the exact checkpoint, runtime version and hardware.