MODELCATALOG

Very large, very sparse mixture-of-experts models under the MIT licence.

DeepSeek · Open-weight

DeepSeek

Very large, very sparse mixture-of-experts models under the MIT licence.

Checked as of 2026-10-23

At a glance

Licence
Open source MIT

MIT for code and weights on every listed model, with no use-based restrictions.

Access
APIBedrockHF
Hugging Face
huggingface.co/deepseek-ai
Source
https://huggingface.co/deepseek-ai

Variants 5

Model Params Context Hardware tierGated InputsStatus
DeepSeek-V4.1-Flash deepseek-ai/DeepSeek-V4.1-Flash Card: 552B backbone plus 196B 'Engram' memory; total here is the safetensors count. 8B active in prefill, 16B in decode. 763B16B active 1.04858M Multi-GPU ml.p5.48xlarge and up No text, image current
DeepSeek-V4-Pro deepseek-ai/DeepSeek-V4-Pro-0813 Official 0813 release. Safetensors total is 1.65T. 1600B49B active 1.04858M Multi-GPU ml.p5.48xlarge and up No text current
DeepSeek-V4-Flash deepseek-ai/DeepSeek-V4-Flash-0731 Official 0731 release. Safetensors total is 304B. 284B13B active 1.04858M Multi-GPU ml.p5.48xlarge and up No text current
DeepSeek-V4-Flash-Vision-Exp deepseek-ai/DeepSeek-V4-Flash-Vision-Exp Experimental. Active parameters not stated; context from config.json. 305B 1.04858M Multi-GPU ml.p5.48xlarge and up No text, image current
DeepSeek-V3.2 deepseek-ai/DeepSeek-V3.2 December 2025. On Bedrock. Size from safetensors, context from config.json. 685B 163.84K Multi-GPU ml.p5.48xlarge and up No text legacy

"—" means the figure could not be verified from a primary source on 2026-10-23 and has been left blank rather than guessed. Hardware tier is from total parameters (all experts for mixture-of-experts models), since every weight must sit in memory. SageMaker rates on the open-weight page.

Strengths

  • MIT-licensed and ungated, even the 1.6T-parameter V4-Pro.
  • A 1M-token context across the V4 family.
  • Very sparse: V4-Flash activates 13B of its 284B parameters per token.
  • Frequent checkpoint updates (V4.1-Flash-1017 on 2026-10-17).

Watch out for

  • Every V4 model is multi-node territory at full precision. Sparse activation saves compute, not memory.
  • No Jinja chat template is shipped, so standard chat tooling needs adapting.
  • Bedrock carries V3.2, V3.1 and R1, not V4.
  • Card parameter figures differ from the HF safetensors totals for some models.