DeepSeek's first V4 vision model is MIT-licensed, and 307 GB

DeepSeek put its first multimodal model in the V4 family on Hugging Face on 31 August, under an MIT licence, and was still updating the repository the following morning. It is called DeepSeek-V4-Flash-Vision-Exp, the Exp is doing real work in that name, and it had been downloaded close to 18,000 times within a day of appearing.

The coverage framed it as an open-source vision model that competes with the frontier labs. Both halves of that are true in a narrow sense and misleading in a practical one, and you can check every correction below against DeepSeek's own files rather than anyone's write-up.

Open weights, 307 gigabytes#

The model card does not print a parameter count, but Hugging Face publishes a census of what is actually in the weight files. Adding it up: 304.6 billion parameters, stored as roughly 296 billion in 8-bit integers, 6.3 billion in FP8, and about 2 billion in bfloat16. That comes to around 307 GB you would have to download and then hold somewhere your accelerator can reach.

The configuration is a mixture of experts: 256 routed experts with 6 active per token, across 43 layers. That design keeps the compute per token far below what 304 billion parameters suggests, which is the entire point of the architecture. It does not shrink the download, and it does not let you leave the other 250 experts on disk, because any token might route to any of them.

It also ships already quantized. The config declares FP8 in the e4m3 format with dynamic activation scaling and 128-wide weight blocks. That matters because quantization is normally the lever you reach for when a model will not fit. Here it has been pulled before you arrive.

So: genuinely open, genuinely permissive, and not a model you will be running on a workstation. If what you want is image understanding on your own hardware, the honest route is a small open vision model served through something like Ollama, which is a different class of thing and will actually start. Treat this release as news about where open weights are going, not as something to download tonight.

Every benchmark is self-reported, and the platform says so#

DeepSeek publishes a table of eleven benchmark rows comparing the new model against its own DeepSeek-V4-Flash-0731 and against Opus-4.8. The numbers are plausible and the methodology is stated, which is more than many releases manage: the text agent benchmarks use the minimal mode of DeepSeek Harness, at maximum reasoning effort, with temperature 1.0 and top_p 0.95.

They are also entirely self-reported, and Hugging Face's metadata is explicit about it. Three of the results are indexed on the platform, and each one carries the field "verified": false with its source recorded as the model card itself. That is not an accusation of anything. It is the difference between a number a lab published and a number a third party reproduced, and the platform tracks that distinction precisely so readers can see it.

There is no independent reproduction of any of these figures yet. For a model released within the last two days, there would not be.

Two of the multimodal wins are against a model that cannot see#

The most striking rows in the table are the multimodal ones, where the new model appears to leap over its predecessor: ApexBench goes from 26.2 to 36.5, and Agents' Last Exam from 25.2 to 27.3.

DeepSeek marks both of those baseline numbers with a dagger, and the footnote explains it: on those two benchmarks, DeepSeek-V4-Flash-0731 "ignores the multimodal elements in the input". The older model is text-only. Those two rows are not measuring a better vision model against a worse one, they are measuring a vision model against a model that was handed images and could not look at them.

That is exactly the sort of caveat that survives in a model card and evaporates in a headline. DeepSeek disclosed it. It is worth reading the footnotes on any comparison table, including this one.

Against Opus-4.8, it leads on three rows out of eleven#

Counted rather than characterised, using DeepSeek's own table: the new model comes out ahead of Opus-4.8 on DeepSWE (59.3 against 58.0), Agents' Last Exam (27.3 against 25.7), and ZeroBench at pass@5 (35.0 against 34.0). It trails on the other eight, including by twelve points on NL2Repo and eight on DSBench-Hard, and it loses Chartography and Toolathlon by margins under a point.

Three of eleven, with several of the losses narrow, is a genuinely respectable showing for a model you can download under an MIT licence. It is not the same claim as beating a frontier model, and the table is right there.

What this actually tells you#

The interesting fact is not the benchmark scores. It is that an experimental multimodal model at this scale shipped under MIT, with its methodology stated and its awkward footnote left in, two days after the weights were built.

That is a real continuation of what open-weight releases have already changed about who controls the stack. It just arrives with a practical asterisk that keeps getting dropped: the licence tells you what you are allowed to do, and the parameter census tells you what you are able to do. Those are different questions, and for most readers the second one is the binding constraint.

If you want to follow this properly, watch for the first independent evaluation. When somebody outside DeepSeek reproduces one of these rows, that number becomes worth something it is not worth today.