DeepSeek V4.1 Flash vs V4 Flash 0731: what changed, and what it costs
Two models called Flash, six point seven times apart in price. One reads images and reasons every turn, the other is the cheap workhorse. Which one for which job.
DeepSeek V4.1 Flash went into the LU Labs Cloud catalog on 10 September 2026, the day the weights went public. It sits in the picker next to DeepSeek V4 Flash 0731, which has been there since August, and the shared word in the name is doing a lot of damage already.
They are not two builds of the same thing. They are a generation apart, and the price difference between them is larger than the price difference between most pairs in the catalog.
The short version
Keep using V4 Flash 0731 for volume work. Reach for V4.1 Flash when you need image input, or when the job is a long agent run rather than a single answer.
Side by side
| V4 Flash 0731 | V4.1 Flash | |
|---|---|---|
| Backbone parameters | not published | 552 billion |
| Active per token | not published | 8 billion in prefill, 16 billion in decode |
| Context window | 1,048,576 | 1,048,576 |
| Reads images | No | Yes |
| Thinking | Think button, and it works | Always on |
| Credits per 1M output | 18,000 | 120,000 |
That last row is the one that decides most projects. One million output tokens on V4 Flash 0731 costs 18,000 credits. The same million on V4.1 Flash costs 120,000. Six point seven times, for work that does not need what the newer model added.
What the new one actually added
The architecture change is about memory, not intelligence. Every token you send a model leaves a key and value entry behind that has to stay in memory for the rest of the conversation. On a long agent run that cache, not the weights, is what grows all day.
DeepSeek states the global cache for V4.1 Flash at 890 bytes per token. On a full million token context that works out to roughly 930 MB. Their own comparison puts it at about a quarter of what V4 Flash needed.
That is why their published benchmark table has the shape it has. On agentic tests, the long horizon ones where the context keeps growing, V4.1 Flash takes the top row repeatedly, including against models with far more active compute. On single shot reasoning questions it does not, and it scores slightly below its own V4 Pro sibling on GPQA Diamond. Those are DeepSeek's numbers, run at their maximum reasoning setting, and they should be read the way any vendor table is read.
The second addition is image input. V4 Flash 0731 returns an error when you attach one. V4.1 Flash reads it. We checked with three small images of different colours before putting it in the picker, and it named each one correctly.
Why the Think button is locked on
Every other reasoning model in the catalog gives you the switch. This one does not, and the reason is a measurement rather than a policy.
DeepSeek built the model with a reasoning effort dial from 1 to 100, not an on and off switch. When an off signal reaches it through the usual API convention, something unhelpful happens. We ran the same question three times each way:
- Left alone: 42 output tokens, about 120 characters of separate reasoning, correct answer, every run.
- Told to stop reasoning: the separate reasoning field came back empty and the turn grew to 51, 53 and 82 output tokens.
The thinking did not stop. It stopped being labelled, and the monologue moved into the answer. A switch that makes your reply longer, dearer and messier is worse than no switch, so the catalog marks this model as always reasoning and the button stays where it belongs.
What five euros buys on it
A 5 EUR credit pack is 165,000 credits and needs no subscription. On V4.1 Flash that is about 1,375,000 output tokens. On V4 Flash 0731 the same pack goes roughly six point seven times further, which is the whole argument for keeping the older one as the default.
For comparison inside the same picker, Kimi K3 costs 1,425,000 credits per million output tokens. V4.1 Flash reads images too and comes in at about a twelfth of that.
Where it sits in the catalog
Every plan and every credit pack carries the whole chat catalog, so there is no tier to buy first. The choice is only ever price against the job.
- V4 Flash 0731 for bulk drafting, summarising, classification and anything you run thousands of times.
- V4.1 Flash when the input includes images, or when an agent is going to spend two hundred steps in the same conversation.
- V4 Pro 0813 when you want the strongest single answer from this family and the volume is small.
Using it
Three routes, same model behind all of them. Open the Studio at lu-labs.ai and pick it from the model list in Chat. Or flip the Cloud switch in the desktop app, which pulls the catalog from the server, so no update is needed for a model added today. Or create a personal API key in account settings and point any OpenAI compatible client at https://lu-labs.ai/api/inference/v1 with the model id deepseek-ai/DeepSeek-V4.1-Flash.
The weights are MIT licensed and public, so running it yourself is allowed. It is also 510 GB of original checkpoint, and the smallest community quant published on day one is 168.9 GB, which is a different conversation entirely.