DeepSeek V4.1 Flash hands-on tests and comparisons

A user spots deepseek-v4.1-flash:1017 callable on Ollama Cloud, noting the model is labeled vision-capable yet vision is actually disabled.

Hands-on: with the same prompt and one-shot run, the 1017 build produced a playable endless underwater game with obstacles, creatures, scoring and better water effects, while the September V4.1-Flash build struggled even to render rays and was never posted.

A hands-on complaint: no voice input, unaware of ChatGPT Astra or V4.1-Flash-1017, no projects or skills, unfit for work, and admits privacy leakage risk.

The poster says 1017 nears Astra on some coding/agent tasks at a fraction of the price, and asks for hands-on comparisons on reasoning, tool use and long-horizon work.

The author says V4.1-Flash-1017 produced a Blender piece with only dsh and its Blender plugin and no supplied assets, costing under 10 RMB, and is struck that this is a 700B mid-tier model.

A practical walkthrough of DeepSeek V4.1 Flash 1017 using a CSV cleaner exercise, including full prompts, five-row acceptance data, an error-fix template, Search/Fetch usage and cost math.

A user reports V4.1 Flash 1017 ranks first on their internal benchmarks, especially security hunting, with low hallucinations, high perception, speed and low cost.

Oran Ge tried making video with DSH and V4.1-Flash-1017; after 50 minutes he called the result 'garbage' — scripts run and shots move, but it looks obviously AI-made.

The official release puts Flash-1017 on web, app and API with DeepSWE up from 74.2 to 77.0 and Terminal-Bench from 31.2 to 42.3, keeping the deepseek-flash name and pricing.

A user's collection lists a ~748B GGUF quant, a ~763B safety-guardrail-ablated variant and its GGUF version, framed as an image-text-to-text model.

On BuildingBench's 12 building tasks, where models reconstruct a castle from four photos, DeepSeek V4.1 Flash 1017 ranks in the top 5 of all models; the eval code is open source.

The official Hugging Face page lists V4.1-Flash-1017 gains in reasoning, coding and agentic tasks: HLE 42.6, Terminal-Bench 42.3, DeepSWE v1.1 77.0, ProgramBench 29.1, with lower execution time and cost.