DeepSeek V4 Flash Vision Weights Are Out, Under MIT

This is a usable release, not an announcement. The weights for DeepSeek-V4-Flash-Vision-Exp — DeepSeek’s first multimodal model in the V4 family — landed on Hugging Face on August 31, 2026 under a plain MIT licence, confirmed on the repository’s own licence field and in the raw LICENSE file, whose first line reads “MIT License.” Ten days earlier the model existed only behind DeepSeek’s hosted API.

The unusual part is not the licence — MIT is DeepSeek’s established posture — but what sits next to the weights. An hour after the 48-shard checkpoint finished uploading, DeepSeek committed a documented reference implementation: standalone prompt-encoding code, a readable PyTorch inference stack covering the vision encoder, and a conversion script. Most open-weight drops ship safetensors — the file format the trained weights are stored in — and a card. This one ships the working drawings too.

First, terms. “Open weights” means the lab publishes the trained model files so anyone can download and run them. On August 21, 2026, DeepSeek-V4-Flash-Vision-Exp had none of that: the model launched API-only, with image-token billing we covered at the time — and our own report stated plainly that no open weights, licence, or Hugging Face repo existed for it. That sentence stopped being true on August 31.

The repository’s commit history pins the sequence to the hour. The repo was created at 06:16 UTC on August 31. Forty-one minutes later, at 06:57, the weights themselves finished uploading — 48 safetensors shards that Hugging Face’s own metadata parser totals at 304,646,824,126 parameters, roughly 304.6B. At 07:57 came the commit that makes this release different: the reference implementation. Follow-up commits through midday August 31 added README material, a thinking-mode encoding example, and evaluation-results files.

One detail worth recording: DeepSeek’s own API changelog — which carries a dated entry for the August 21 API launch — shows no separate entry for the weights release, and we found no announcement on DeepSeek’s homepage either. Those are the two official surfaces we checked; we cannot rule out a social post elsewhere. On the record DeepSeek itself keeps, the Hugging Face commit is the announcement — and no reason for the ten-day gap appears anywhere on it. Where a vendor does state a reason for holding weights back, that statement carries information of its own; DeepSeek offered none.

The licence claim in this post’s title rests on two first-party artifacts, not on coverage. The Hugging Face model card’s machine-readable licence field for deepseek-ai/DeepSeek-V4-Flash-Vision-Exp reads mit. And the raw LICENSE file in the repository is the plain MIT text — no appendix, no acceptable-use section, no revenue threshold, no attribution mandate. The distinction between reading the file and trusting the badge is not pedantry: licence-field metadata elsewhere has been wrong about exactly this question, which is why we check the text itself.

Plain MIT means the full stack of rights with no carve-outs: use, copy, modify, merge, publish, distribute, sublicense, sell. Fine-tuning, derivative checkpoints, commercial self-hosting, resale — none of it gated by who you are or how much revenue you book. DeepSeek attached it without ceremony.

The model itself, per the card and the repo’s own config.json: a mixture-of-experts transformer built on the DeepSeek-V4-Flash architecture with added visual modules and continued training — 256 routed experts, 43 hidden layers, a 1,048,576-token context window unchanged from the text-only V4 Flash, and a dedicated 32-layer vision tower. One number DeepSeek does not publish for this checkpoint is an active-parameters-per-token count, so we do not print one; figures circulating for “V4 Flash” trace to the earlier text-only model, not this release.

A definition, because the distinction carries this section: the weights are the trained numbers — necessary, but inert without code that knows how to run them. A reference implementation is that code, written to be read: it shows, executably, how prompts are encoded, how images become tokens, and how the forward pass is wired. Most labs leave that work to the community or to inference frameworks. DeepSeek shipped it in the repo, and its own inference/README.md is precise about intent: “This directory contains a readable reference implementation rather than a production serving engine.”

We verified each claimed component against the repository’s actual file listing rather than taking the README’s word for it. The inference/ directory carries the vision encoder and aligner (vision.py), the model definition with the architecture components the README names — DFlash attention, MoE, and Hyper-Connections — in model.py, kernels (kernel.py), generation with the DSpark speculative forward path (generate.py), and a checkpoint-conversion script (convert.py). The encoding/ directory holds encoding_dsv4.py, which DeepSeek documents as the standalone prompt-format reference: “It supports multi-turn conversations, tool calls, thinking modes, and interleaved image content blocks without importing the inference implementation.” DSpark is not marketing language either — the config file carries its block size, target layers, and Markov rank as concrete fields.

One small cross-check ties the released config to the hosted product. The config’s vision_max_n_token field reads 384 — the same number the API’s image billing cap stated at the August 21 launch. Two independent artifacts, the hosted API’s pricing page and the released checkpoint’s config, agree on how large an image can get in tokens. That is the kind of consistency you can only check because the weights are out.

On capability, the model card publishes a benchmark table — all figures vendor-stated, not independently reproduced. Against the text-only V4-Flash-0731 and Anthropic’s Opus 4.8, DeepSeek reports Terminal Bench 2.1 at 83.9 versus 82.7 and 85.0 respectively, and DeepSWE at 59.3 versus 54.4 and 58.0 — on that second benchmark, DeepSeek’s card shows its own model ahead of both comparators.

On August 28 we published a post whose title has the same shape as this one: “GLM-5.3’s Weights Are Out. The Licence Is Not MIT”. Z.ai had just shipped its 753B flagship’s weights under a bespoke, vendor-named licence — near-MIT in its grant language, but with a revenue-gated security-review clause for the largest Model-as-a-Service operators. Three days later, DeepSeek faced the identical decision for a flagship-class release and made the opposite call: no bespoke document, no gate, no named licence of its own. Plain MIT.

The two releases also resolve differently on timing. The announcement-to-weights gap is a number we track across 2026 releases, and the two gaps are measured from different starting points: ten days from DeepSeek’s API launch to weights, fourteen days from Z.ai’s GLM-5.3 announcement to weights — with one more asymmetry worth naming: Z.ai stated a reason for its hold, a safety-evaluation period; DeepSeek stated nothing at all.

That line, from our own August 21 report, is the loop this post closes. Ten days is short enough that the two posts describe what reads in hindsight as one staged release: hosted first, weights after. But nothing DeepSeek published promised the second stage — which is exactly why “is this a usable release or an announcement?” was a live question until the commit landed.

“Open weights” coverage usually stops at parameter counts and benchmark scores. A more useful axis for anyone who will actually touch the files is what ships beside the weights. The rows below come from the four repositories’ own Hugging Face file listings and licence fields — DeepSeek’s new release, DeepSeek’s previous flagship drop, and two prominent peer releases with weights on the platform.

Two honest readings of that table. Against the wider field, a documented, separately-runnable reference implementation is rare — of the two peer flagships here, one ships nothing beyond the minimum and the other ships library glue. But it is not a first for DeepSeek: the V4-Pro-0813 weights drop in mid-August followed the identical pattern. The accurate framing is that DeepSeek is continuing a house style almost nobody else follows — which, for anyone betting on the V4 line, is arguably better news than a one-off gesture would be.

The release converts into different actions depending on what you do with models. The common thread: because the licence is plain MIT and the repo documents its own internals, the usual first week of open-weight due diligence — licence archaeology, prompt-format reverse-engineering — mostly is not needed here.

The wider pattern this fits: three days apart in late August, the two most watched open-weight labs answered the same licensing question in opposite directions, and the only way to know which answer you got is to read the repository itself. That per-release check — licence file, what ships beside the weights, what the vendor’s own README claims versus what the file listing shows — is exactly the model-adoption discipline our AI transformation engagements build into procurement, routing, and governance workflows.