DeepSeek releases V4.1 Flash, an MIT-licensed multimodal model with an unusual architecture

DeepSeek released V4.1 Flash on 10 September 2026, with the weights published on Hugging Face under the MIT licence. It is a 552-billion-parameter multimodal mixture-of-experts that takes images and text as input and produces text, carries a 1-million-token context window, and was trained on a reported 45 trillion tokens of mixed text and images. The MIT release, read at source, carries no revenue threshold and no territory restriction, which puts it among the most straightforwardly usable of the large open models.

An architecture that stands out

Two things make this more than another large release. The first is how little of it runs at once. Of the 552 billion parameters, the model activates roughly 8 billion per token while reading the prompt and roughly 16 billion per token while generating the reply. That asymmetry is deliberate: reading a long input is made as cheap as possible, and a little more capacity is spent on writing the output. It means the model computes far more like a small one than its total size suggests, which is what makes a model this large practical to serve at volume.

The second is the shape. DeepSeek describes a Causal Encoder-Decoder design, a 20-layer causal encoder followed by a 20-layer decoder, rather than the decoder-only architecture that almost every current large language model uses. Encoder-decoder designs were common in an earlier era of the field and then fell out of favour for generative models; seeing a leading open lab return to one, and pair it with native image understanding, is genuinely interesting. As with any new design, the honest position is to wait for independent testing before drawing conclusions about how it performs in practice.

Who it is for

For local use, be clear about scale. A 552-billion-parameter model is server-class: even at a 4-bit quantisation the weights alone need on the order of 280GB of memory, and at the released precision considerably more, so this is a multi-GPU server or a high-memory machine rather than a consumer card. That is an estimate, and the exact figure depends on the quantisation and serving stack. Production GPU servers such as vLLM and SGLang are the realistic route to running it. DeepSeek also offers it as a hosted API, with off-peak pricing the company lists at 0.15 US dollars per million input tokens and 0.60 per million output tokens (DeepSeek, September 2026), which is the relevant comparison for anyone weighing self-hosting against renting.

The audience that benefits most is anyone who needs multimodal understanding at high volume and wants to keep it in-house under a permissive licence, or who is renting inference and wants a cheaper, openly licensed option than a closed multimodal API.

Where it sits

This continues DeepSeek’s consistent posture of shipping capable models under the MIT licence, following DeepSeek V4 in April and the earlier DeepSeek R1. That consistency is worth noting against the wider picture we have been tracking: some strong open models now carry bespoke licences with revenue gates, and some carry territory clauses that exclude UK users outright. DeepSeek continuing with plain MIT, on a multimodal model at that, is a point in the open ecosystem’s favour. It also sits among a run of very large open mixture-of-experts models, alongside the likes of Kimi K3 and the GLM-5.3 line, where the story is increasingly about how few parameters actually compute per token rather than how many the model has in total, from Kimi K3 to Thinking Machines Lab’s Inkling and the GLM-5.3 line.

We have added DeepSeek V4.1 Flash to the model catalogue with its confirmed licence and hardware requirements, and we will fill in independent benchmark results as they appear.