DeepSeek V4.1 Flash 1017: What Changed, and What It Means in LU Cloud
DeepSeek published V4.1 Flash 1017 on 17 October 2026. Since 18 October, 21:54 UTC, it is the model behind the DeepSeek V4.1 Flash entry in LU Cloud, and the picker shows it as DeepSeek V4.1 Flash 1017.
You do not have to do anything. If you picked V4.1 Flash in a chat, saved it as your default or call it over the API with deepseek-ai/DeepSeek-V4.1-Flash, those requests now go to 1017. The old id keeps working.
What stays the same
DeepSeek calls 1017 a minor update to V4.1 Flash. The checkpoint on Hugging Face is the same size, 510.3 GB, under the same MIT licence. In LU Cloud it keeps what the old entry had:
- 1,048,576 tokens of context
- image input
- native tool calls, so Agent mode and the coding agent run on it as before
- reasoning on every turn, with the Think switch locked on
Before the switch we ran the new version next to the old one: a tool call, an image to describe, and the same reasoning prompts. Both behaved the same way, so none of the settings changed.
What DeepSeek says is better
The release note says 1017 scores higher on benchmarks, finishes tasks faster and writes shorter outputs than the version before it. The model card adds a table. These are DeepSeek's numbers, run by DeepSeek, comparing 1017 with the V4.1 Flash it replaces:
| Benchmark | V4.1 Flash | V4.1 Flash 1017 |
|---|---|---|
| HLE without tools | 36.8 | 42.6 |
| Terminal-Bench 4.0 | 31.2 | 42.3 |
| DeepSWE v1.1 | 74.2 | 77.0 |
| ProgramBench | 20.3 | 29.1 |
| AutomationBench | 54.8 | 60.5 |
| Agents' Last Exam | 31.8 | 35.4 |
| Apex Bench v1.1 | 64.6 | 72.1 |
| Chartography, with tools | 78.9 | 83.0 |
| BabyVision, with tools | 89.6 | 92.5 |
| ZeroBench, with tools | 49.0 | 51.0 |
The largest gains are on the long agent tests, Terminal-Bench 4.0 and ProgramBench. The three image tests move by two to four points.
We have not measured output length ourselves yet. If DeepSeek's claim holds, the same task costs fewer output tokens on 1017 than it did on V4.1 Flash. For API users, the model card now recommends temperature 1.0 and top_p 1.0.
What it costs in LU Cloud
The catalog price went down with the switch. DeepSeek V4.1 Flash 1017 draws 20,000 credits per million input tokens and 60,000 per million output tokens. Before the switch the catalog listed V4.1 Flash at 30,000 and 120,000.
DeepSeek V4 Flash 0731 is still the cheaper choice for bulk work. It draws 18,000 credits per million output tokens, about a third of what 1017 does, but it cannot read images.
Every plan and every credit pack carries the whole chat catalog, so there is no tier to buy for this model. Plans and packs are on the pricing page.
Using it
In Locally Uncensored, flip the Cloud switch and pick DeepSeek V4.1 Flash 1017 in Chat. The app pulls the cloud catalog from the server, so the new entry is there without an app update. In a browser, the same model is in the Studio at lu-labs.ai.
From your own code, create a personal API key in account settings, point any OpenAI-compatible client at https://lu-labs.ai/api/inference/v1 and use the model id deepseek-ai/DeepSeek-V4.1-Flash-1017.
Can you run 1017 locally?
The weights are public under MIT, so you are allowed to. The hardware math has not changed: the checkpoint is the same 510.3 GB as before, which puts it out of reach for a desktop machine. The numbers are in Can You Run DeepSeek V4.1 Flash Locally? and the architecture is explained in DeepSeek V4.1 Flash Explained. If you would rather run a model that fits your own card, with nothing leaving your machine, that is what Locally Uncensored is for.
In the handbook
- All DeepSeek guides, every generation on one page
- LU Cloud, the account, the Cloud switch and the two marks in the picker
- Chat, the model picker, the context window and thinking mode