DeepSeek V4.1 Flash Xploded is on the line
An uncensored DeepSeek V4.1 Flash joins Xploded: a million tokens of context, tool calling, adjustable reasoning and the longest replies on the line. What it is, how it was tested and where it fits.
What it is
DeepSeek V4.1 Flash Xploded (deepseek-v4.1-flash-xploded) is an uncensored build of DeepSeek V4.1 Flash, the model DeepSeek released on 10 September 2026 as the first of a new architecture family. It reads a million tokens, calls tools, reasons before it answers and writes up to 393,216 tokens in a single reply, the longest of any model on the line. It is text only.
The test it passed
Every Xploded model has to do two things at once: answer without refusing, and call tools correctly. On 18 October 2026 we sent it the same request every model on the line has passed, a get_weather tool and a question that cannot be answered without it. It came back with finish_reason: "tool_calls" and {"city": "São Paulo"}. The whole bar is in how a model gets into Xploded.
Where it fits
Artificial Analysis scores the base DeepSeek V4.1 Flash at 39 on its Intelligence Index, between GLM 5.3 Flash at 42 and MiMo V2.6 Flash at 38. Uncensored builds are not benchmarked separately, so read that as a guide to the base model.
Three things set it apart on the line. It writes the longest replies, three times GLM 5.3's limit. Its reasoning effort can be set per request, or switched off for bulk work, where the GLM models always think. And it costs little to re-read a long history, which is most of what an agent sends. It does not read images, so screenshots belong with GLM 5.3 Flash or MiMo. The head-to-head pages have the detail: against GLM 5.3 Flash and against MiMo V2.6 Flash.
About 1017
DeepSeek shipped an update to V4.1 Flash on 17 October 2026, called 1017, with better agent and coding scores and shorter answers. Its benchmark table is DeepSeek's measurement of DeepSeek's own checkpoint, and we do not claim those gains for the uncensored build. What 1017 changed has the full table.
How to use it
It sits behind the same OpenAI-compatible and Anthropic-compatible endpoints as the rest of the line, at https://api.magmarouter.com, with the same key. Name it by its id and you are done. The API quickstart has curl, Python and JavaScript, Codex CLI and the Claude Code guide cover the agents, and you can try it in chat first.