DeepSeek V4.1 Flash is out: 552B, MIT, and nothing can run it yet

Description and use

<p>DeepSeek published <strong>DeepSeek-V4.1-Flash</strong> on Hugging Face on <strong>10 September 2026 at 02:17 UTC</strong>, MIT licensed and ungated. I spent the day reading the files and poking the model through an OpenAI compatible endpoint instead of reading the announcement. Here is what is worth knowing before you start a download measured in hundreds of gigabytes.</p> <h2> TL;DR </h2> <ul> <li> <strong>552B backbone</strong>, 8B active in prefill and 16B in decode, 1,048,576 token context, reads images.</li> <li>Original checkpoint: <strong>510.3 GB</strong> across 48 shards, FP8 with

Problem

<p>DeepSeek published <strong>DeepSeek-V4.1-Flash</strong> on Hugging Face on <strong>10 September 2026 at 02:17 UTC</strong>, MIT licensed and ungated. I spent the day reading the files and poking the model through an OpenAI compatible endpoint instead of reading the announcement. Here is what is worth knowing before you start a download measured in hundreds of gigabytes.</p> <h2> TL;DR </h2> <ul> <li> <strong>552B backbone</strong>, 8B active in prefill and 16B in decode, 1,048,576 token context, reads images.</li> <li>Original checkpoint: <strong>510.3 GB</strong> across 48 shards, FP8 with

Outcome

<p>DeepSeek published <strong>DeepSeek-V4.1-Flash</strong> on Hugging Face on <strong>10 September 2026 at 02:17 UTC</strong>, MIT licensed and ungated. I spent the day reading the files and poking the model through an OpenAI compatible endpoint instead of reading the announcement. Here is what is worth knowing before you start a download measured in hundreds of gigabytes.</p> <h2> TL;DR </h2> <ul> <li> <strong>552B backbone</strong>, 8B active in prefill and 16B in decode, 1,048,576 token context, reads images.</li> <li>Original checkpoint: <strong>510.3 GB</strong> across 48 shards, FP8 with

Evidence

DeepSeek V4.1 Flash is out: 552B, MIT, and nothing can run it yet

Source