Last time DeepSeek’s Flash line appeared on this page, the headline was how small it could get. A 284-billion-parameter model, squeezed to roughly 103GB, ran on a well-equipped Mac. The new Flash is a different animal, and pretending otherwise would be like reviewing a delivery van as if it were still the scooter.
DeepSeek-V4.1-Flash arrived on September 10, 2026 as the smallest member of a new architecture family, with native image understanding. On October 17 DeepSeek released the 1017 update, which it calls minor: better reasoning, coding and visual scores, shorter outputs, faster tasks. The weights are public under the MIT license. The stored checkpoint is about 763 billion parameters, according to the Hugging Face file sizes, and the launch coverage we found gives no official quantized sizes. DeepSeek’s own advice is blunt: practical speed takes hundreds of gigabytes of VRAM or unified memory.
“Flash” still means something real, but not small. Artificial Analysis lists the September base at 552 billion parameters with about 16 billion active per token. In a mixture-of-experts model, a router picks a few specialist sub-networks for each word, so each step is cheap. The whole library still has to sit on the shelf. A 16B-active model runs quickly if the building is big enough; it does not fit in a smaller building.
What do you get for the hardware? DeepSeek’s launch table has the model at 77.0 on DeepSWE v1.1, 60.5 on Automation-Bench, 35.4 on Agents’ Last Exam and 72.1 on Apex, leading Claude Opus 5, GPT-5.6 Sol and GLM-5.3 on each of those. It trails Opus 5 on Terminal-Bench 4.0, ProgramBench, NL2Repo and Humanity’s Last Exam. All of it is DeepSeek-run. A vendor choosing its own harness is like a student grading their own exam: the effort may be honest, and you still want a second marker.
The second marker has only partly arrived. Artificial Analysis measured the September checkpoint at 39 on its Intelligence Index, three points under GLM-5.3-Flash on the same version of the index. The 1017 update has not been measured independently as of this review, so the practical expectation is a model that is better at agent work than the September one and not yet shown to lead on broad intelligence.
That makes the fit narrow but clear. If you run a private GPU cluster, want an MIT license without custom clauses, need a model that reads images and long documents, and have the engineering time to test a quantized build against your own tasks, this is a serious option. If you have a single workstation, look at Qwen3.8-27B first. If you simply want cheap tokens and privacy is not the point, the hosted API at $0.15–$0.30 input and $0.60–$1.20 output per million tokens, depending on the hour, is the easier door, with the caveat that your data goes to DeepSeek.
The honest summary: a stronger model than the one it replaces, and a harder one to own. Our score falls because Local / Private AI asks who can actually run the thing, and the answer for this checkpoint is fewer people than before.
