DeepSeek V4.1 Flash hands-on tests and comparisons

A user built 3D worlds and an FPS game with DeepSeek V4.1 Flash, judging it no weaker than Opus 5.5 and confirming the official claim that the Harness is optimized for game building.

The post notes 1017 is live on app, web and API, with Terminal-Bench up from 31.2 to 42.3, DeepSWE at 77.0 versus Opus 5's 74.0, and open weights under MIT.

Hands-on speed test: V4.1 Flash 1017 outputs about 274 tok/s in DeepSeek Harness High mode, versus just over 30 tok/s for GPT-6.1 Sol.

DeepSeek released V4.1-Flash-1017, scoring 77.0 on the DeepSWE software engineering benchmark and opening its Search and Fetch APIs so one key covers inference plus live web retrieval.

Both outputs impressed on the same prompt, with Astra winning on realism, but the author argues V4.1 Flash 1017 already wins by competing at all, given its price and speed.

Asked to review a repo already vetted by GPT-5.6 Sol, 1017 quickly found several bugs and hidden risks and fixed them all, with the user praising it as fast, cheap and more usable than Sol.

A video pits the new DeepSeek V4.1 Flash 1017 build against the 0910 release, though the post itself gives no concrete results.

A user explains that dated sub-versions under an approved v4.1 need no new clearance, while a new v4.2 would require fresh government approval.

On one shared prompt Opus 5.5 shows better polish and detail, but DeepSeek takes about an hour at $0.23 versus roughly $6 for Opus, about 1/26 the cost, which the author calls impressively close.

A user posts video claiming DeepSeek V4.1 Flash beats Argon, the Gemini 4 model Google still gates for most people, and questions how a Flash-tier model can win.

Breakdown of official numbers: DeepSWE Pass@1 hits 77.0 at about $0.2 per task versus Opus 5's ~74 at over $10, while HLE and Terminal-Bench still trail Opus; Search/Fetch API costs $0.75 per 1,000 calls.

A video test of DeepSeek V4.1 Flash on 460M tokens for about 24 yuan finds its 3D scenes reach roughly 80% of Gemini 4 Argon's quality.