PyyanThe AI Index

Pyyan / Compare / DeepSeek V4.1 Flash vs Atria Dawn Preview

DeepSeek V4.1 Flash vs Atria Dawn Preview

2 of 5

Open-Weight Models · verified 16 Sept 2026

DeepSeek V4.1 FlashDeepSeekcurrent
SAtria Dawn PreviewShanghai AI Laboratorycurrent
3 slots left
SpecificationDeepSeek V4.1 FlashAtria Dawn Preview
SummaryTwice the parameters of the Flash it replaces, and cheaper to run.744B agentic weights under the MIT licence, published without a launch.
Context1M256K
Input ($/Mtok)$0.3Not published
Output ($/Mtok)$1.2Not published
LicenceMITMIT
Max output256KNot published
Parameters552B total, 8B / 16B active744B total
Released10 Sep 202614 Sep 2026
Model IDdeepseek-flashinternlm/Atria-Dawn-Preview
CategoryOpen-Weight ModelsOpen-Weight Models
OfficialDeepSeek ↗Shanghai AI Laboratory ↗

Highlighted rows are where these differ.

DeepSeek V4.1 Flash

  • 552B total, 8B active while reading the prompt and 16B while generating, with a 1M token context
  • DeepSeek says it beats V4-Pro; V4-Pro requests route to it from 14 September 2026
  • MIT licence, native image input, and a KV cache about a quarter the size of V4-Flash's

Best for cheap agentic coding with images.

Full spec sheet →

Atria Dawn Preview

  • Built on Zhipu's open GLM-5.2 rather than trained from scratch
  • 96.0 on DeepSearchQA and 86.5 on CyberGym, both reported by the lab
  • BF16 and FP8 checkpoints, served through SGLang or vLLM

Best for agentic work on your own cluster.

Full spec sheet →