deepseek-v4.1POST

DeepSeek V4.1 Flash 1017 open weights and API pricing clarified

The post cites specifics: MIT open weights around 510 GB, a 552B backbone plus 196B Engram activating only 8B (prefill) / 16B (decode) per token, API at $0.15/M input and $0.60/M output off-peak, with quantized builds running on high-RAM machines or a few GPUs.

Goodsmash.eth

@Web3IsWildArchived 11 days agoOriginally posted 2026.10.21

Goodsmash.eth @Web3IsWildOriginally posted 2026.10.21

Open original on X

You clearly never looked at the Hugging Face page. DeepSeek-V4.1-Flash-1017 is MIT open weights: ~510 GB download, 552B backbone + 196B Engram, but only activates 8B (prefill) / 16B (decode) per token. API is $0.15/M input and $0.60/M output off-peak. Quantized builds run on high-RAM machines or a few GPUs, not a $100k setup. Inference ≠ training. The model card and pricing page are public.

x.com/web3iswild/status/2112740802346524907Archived 2026.10.21 12:08