DeepSeek V4.1 Flash hands-on tests and comparisons

Cost breakdown: for 130 yuan, a Claude Opus 5.5 subscription yields ~2.9B tokens per month versus ~2.12B off-peak tokens for DeepSeek 1017.

The author used DeepSeek 1017 via Workbuddy's free routing to replicate a video: output came fast, but repeated revisions got messier and long-horizon self-checking was weak.

A user notes only 6 of 15 benchmarks list a DeepSeek V4.1 Flash 1017 score, and it beats M3.1 Flash on 4 of them (DeepSWE, Terminal Bench, NL2Repo, BabyVision), questioning why the rest were not run.

After hands-on use, the poster says other open models feel fake next to Opus 5.5, and DeepSeek's latest Flash 1017 is the first usable one at a fraction of the cost.

On one identical prompt, DeepSeek V4.1 Flash took 57m 21s for $0.32 while Opus 5.5 took 13m 20s for $2.27, roughly a 7x cost gap.

Hands-on: DeepSeek V4.1 Flash rebuilds a The Last of Us Part 2 scene, getting the boat, motor, rope and water reflections nearly right despite being a Flash model.

A user had DeepSeek V4.1 Flash 1017 generate a 30-second themed short film from a single prompt based on an official notice, spending just over one yuan in tokens and calling the result decent.

A user tried the SINGLE-DGX DeepSeek-V4.1-Flash-1017 build and none ran correctly, mostly garbled text and loops, though they still hold hope for projectMAYA.

The author explains NEAR AI's privacy tiers: DeepSeek V4.1 Flash 1017 is labeled Incognito, where routing hides the account but the model still runs at an external provider, unlike the isolated Private TEE tier.

Powertokens promotes DeepSeek V4.1 Flash on its platform, listing a thinking toggle, high/max effort, logprobs, JSON mode, function tools and streaming usage stats.

MiniMax released M3.1 Flash weights (428B total, 23B active, 1M context); its official comparison table shows it behind DeepSeek V4.1 Flash on DeepSWE (73.7 vs 77.0) and Terminal Bench (38.1 vs 42.3) but slightly ahead on Automation-Bench (62.8 vs 60.5) and tool-free HLE (43.4 vs 42.6).

A short Portuguese post reports that after re-tuning the machine, DeepSeek V4.1 Flash 1017 reached a 6k tok/s peak during decode, with a screenshot attached.