DeepSeek V4.1 Flash hands-on tests and comparisons

In DeepSeek Harness, a one-line prompt produced a 15-second beat-synced video: 10m53s model time, 43m18s tool time, 265 tok/s output, 99% cache hit, 18.6M tokens total, judged close to Opus 5.5.

The same prompt asked DeepSeek V4.1 Flash and Opus 5.5 to build a cinematic 3D animation explaining an EV motor, with the author asking whether Opus is in trouble.

The author says V4.1 Flash 1017 scores 77 on DeepSWE, beating GPT-6.1 Sol, Astra and Opus on the same chart at a very low price, with a 3-minute demo.

A user made a small island scene showing spring blooms, autumn leaves and winter snow with DeepSeek V4.1 Flash, costing $0.40 over 2.5 hours.

A user built a 3D Mario game with DeepSeek V4.1-Flash-1017 in DeepSeek Harness, hitting 99.8% cache and 438K output tokens for $0.30 total.

A single DeepSeek Harness run produced a 3D bee resting on a sunflower, with the author praising the modelling detail and lighting.

Bloomberg reports DeepSeek says the new model is faster and leaner than the previous version, with better benchmarks, quicker task completion and shorter answers, now live on web, app and developer platform.

A roundup of hands-on reports: V4.1 Flash scores 34 on a 105-bug repair test versus 41.7 for Opus 5.5 and 21.7 for the prior Flash, runs 230–300 tok/s at $0.15/$0.6 per million tokens, and DSH uses 63% more time and 91% more tokens than Pi.

Local deployment report: running V4.1 Flash 1017 across 4 DGX Sparks with TP4 gives impressive speeds, with an open-source deployment repo linked.

In a Blender scripting test, Flash matched Fable on simple props at 36x lower cost, but dropped constraints and contradicted its own naming in complex and long builds.

Databricks' Yuchen Jin says DeepSeek-V4.1-Flash-1017 scores 77 on DeepSWE v1.1, matching Opus 5 at 83x lower cost, and announces the model will launch on Databricks.

Called a minor update, but DeepSeek's own table shows Terminal-Bench 4.0 jumping from 31.2 to 42.3, with MIT-licensed open weights at $0.15 per million input tokens.