DeepSeek V4.1-Flash-1017 lifts Terminal-Bench from 31.2 to 42.3 at unchanged pricing

The gains come from self-feedback reinforcement learning applied to the existing Flash model, delivering an 11-point Terminal-Bench jump in about five weeks. One developer said it 'passed my threshold for caring.'

The MIT-licensed weights are welcome, but they come as terabyte-scale downloads with an 8-GPU minimum, so most users will rely on the API. HN discussion fixates on cost per task rather than leaderboard placement.

Pricing gaps between DeepSeek, Mistral Large 4 and Cohere Command A+ (around 9x) are driving cost-led model switching.

The launch coincides with the DHS federal ban, which limits uptake among US government users.