DeepSeek-V4.1-Flash

Checkpoint identity

Official identifier

deepseek-ai/DeepSeek-V4.1-FlashDeepSeek-V4.1-Flash official model card (external source) · DeepSeek-V4.1-Flash checkpoint configuration (external source) · DeepSeek-V4.1-Flash MIT license (external source)

Architecture

Multimodal mixture of experts; causal encoder-decoder with compressed sparse attention and conditional memoryDeepSeek-V4.1-Flash official model card (external source) · DeepSeek-V4.1-Flash checkpoint configuration (external source) · DeepSeek-V4.1-Flash MIT license (external source)

Total parameters

Not verified for this recordPublished rounded counts; total stored weights determine the preliminary footprint, including all MoE experts.DeepSeek-V4.1-Flash official model card (external source) · DeepSeek-V4.1-Flash checkpoint configuration (external source) · DeepSeek-V4.1-Flash MIT license (external source)

Active parameters

Not verified for this recordActive parameters are a compute-related count, not the model's full stored weight count.DeepSeek-V4.1-Flash official model card (external source) · DeepSeek-V4.1-Flash checkpoint configuration (external source) · DeepSeek-V4.1-Flash MIT license (external source)

Modalities

text-input, image-input, text-outputDeepSeek-V4.1-Flash official model card (external source) · DeepSeek-V4.1-Flash checkpoint configuration (external source) · DeepSeek-V4.1-Flash MIT license (external source)

Context

1,048,576 tokensNative configuration where directly verified. Extensions, advertised K-labels and configuration differences are preserved in the notes below.DeepSeek-V4.1-Flash official model card (external source) · DeepSeek-V4.1-Flash checkpoint configuration (external source) · DeepSeek-V4.1-Flash MIT license (external source)

Release date

Exact release date not independently established in the reviewed artifactsDeepSeek-V4.1-Flash official model card (external source) · DeepSeek-V4.1-Flash checkpoint configuration (external source) · DeepSeek-V4.1-Flash MIT license (external source)

Download access

Public checkpoint referenceDeepSeek-V4.1-Flash official model card (external source) · DeepSeek-V4.1-Flash checkpoint configuration (external source) · DeepSeek-V4.1-Flash MIT license (external source)

License

MITThe publisher's exact checkpoint LICENSE uses MIT terms. Preserve copyright and permission notices. The checkpoint's license is separate from any hosted API contract, and neither license nor model-card examples establish deployment performance.DeepSeek-V4.1-Flash official model card (external source) · DeepSeek-V4.1-Flash checkpoint configuration (external source) · DeepSeek-V4.1-Flash MIT license (external source)

Theoretical weight-memory footprint

A weight estimate is unavailable without a verified parameter count.

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 declares FP8 quantization with 32x32 blocks, UE8M0 scales and FP4 experts, plus a BF16 general dtype. Weight, activation and cache formats are distinct. A generic all-FP8 or all-four-bit storage estimate would omit those distinctions.

Serving considerations

• The publisher separately reports a 552B backbone and 196B conditional-memory component. Activated work differs by phase: the card reports 8B in prefill and 16B in decode. A single total or active count is left unknown until all checkpoint components are reconciled.
• The text configuration sets 1,048,576 positions with explicit scaling from a shorter original context. Long-context memory cannot be modeled with a plain attention-cache formula without accounting for its causal encoder-decoder, compressed sparse attention and conditional memory.
• The publisher provides a minimal-inference implementation and checkpoint conversion instructions. No vLLM/SGLang support is asserted here from Hugging Face's automatically generated deployment snippets; an exact maintainer-supported engine path remains a separate verification task.
• For a future evaluation, pin the artifact, conversion path, quantization kernels, image processing and phase-sensitive cache settings. Measure complete requests under representative prompts; publisher cache-size claims are not Aguasto measurements.
• The model is a research reference. Aguasto has not run its deployment commands, verified GPU fit, hosted its weights or offered it through an API.

Framework documentation

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