DeepSeek V4.1 Flash hands-on tests and comparisons

Hands-on 3D building reconstruction test: 81.9 overall (up from 72.9), slightly above Fable 5.1's 81.4 and near Sonnet 5.5's 82.1, at about $3.01 median per task — roughly 1/11 and 1/15 of their cost.

Side-by-side test: the paid tier renders tiny animated ocean waves while the free one shows no visible water, with the author joking the savings cost an afternoon of code fixing.

A user summarizes the Saturday release: 763B total parameters, 1M context, MIT-licensed weights, API price unchanged from the prior version, and coding/agent benchmarks near Opus 5 level.

A roundup of official specs: 763B parameters, 1M context, downloadable weights, Terminal-Bench 4.0 up from 31.2 to 42.3, DeepSWE 74.2 to 77.0, and a $0.75 per 1K search API, while HLE still trails Opus 5.

User test: V4.1 Flash-1017 puts 85% of tokens into thinking and wrote no code after 40k tokens, yet scored 91.5% on 251 questions for just $1.8 total.

Alongside V4.1 Flash-1017, DeepSeek opened Search and Fetch APIs priced at just ¥5 per 1,000 requests, several times cheaper than the tens of yuan other platforms charge.

Official release: gains in reasoning, coding, agent and vision benchmarks, shorter faster outputs, MIT-licensed weights on Hugging Face, unchanged API pricing, plus new Search and Fetch APIs.

Experiential Labs lists DeepSeek V4.1 Flash 1017 with a no-card free trial, claiming 1.05M context and up to 393K output; the poster is now testing long context.

A short reaction arguing that the gap between closed frontier models and open weights is shrinking, or that the 'good enough' bar users care about now sits near open-weight level.

A user jokes about spending the equivalent of eight years of Claude subscriptions to self-host, only for the team to get a poor quantization of V4.1 Flash 1017 at 135 tok/s.

The author attributes GPT-6.1 Sol's 150+ tok/s to competitive pressure and a user shift toward DeepSeek V4.1 Flash 1017, an observation without supporting data.

Official demos show a 3D Forbidden City, a racing game and Portal 2 rebuilt in Godot, with DeepSWE rising from 74.2 to 77; web, app and API updated at unchanged pricing under MIT license.