DeepSeek V4 Vision Open-Weights Leak: Vision-Exp Is API-Only, the Final Model Is Expected Open

Two sentences posted on X this week are the most useful thing anyone has said yet about where DeepSeek's V4 vision line is heading. teortaxesTex, a researcher who follows DeepSeek closely, wrote: "I suspect that yes they will not open weights Vision-exp. They have done such things in V2 gen. But they will definitely open weights V4-Vision-Final." None of that is confirmed — it is an informed prediction, not an announcement — but the first half already matches what happened. DeepSeek-V4-Flash-Vision-Exp, the V4 family's first model with native image input, went live on DeepSeek's own API on August 21, 2026 as an API-only release with no downloadable weights. The second half, if it holds, changes what a developer can actually do with the V4 vision line: self-hosting, local fine-tuning, and a real choice between running the model yourself and calling it.

This is a what-we-know-so-far piece, and the sourcing line matters more than usual. The tweet is an insider read, not a leak of files, and it is unverified by design. What is verified is the current state of DeepSeek-V4-Flash-Vision-Exp — released, priced, benchmarked, API-only — and what is unverified is everything about the final model's release. That distinction runs through every section below.

What the leak actually says

The signal is short, so it is worth parsing precisely. "Vision-exp" is DeepSeek-V4-Flash-Vision-Exp, the experimental model DeepSeek shipped on August 21. "V4-Vision-Final" is the researcher's shorthand for the production V4 vision release that DeepSeek has promised but not yet named. The prediction has two parts: DeepSeek will keep the experimental model's weights closed — a pattern teortaxesTex says holds from the V2 generation — and the final model's weights will be opened.

The strongest V2-generation precedent on the record is DeepSeek-VL2, DeepSeek's open vision-language family from December 2024, released under the permissive DeepSeek Model License with MIT-licensed code. It is exactly the "final opens" half of the pattern. The "experimental stays closed" half is the researcher's reading of DeepSeek's behavior, and it is consistent with how the company has handled preview models since — which is what makes the final-model prediction worth taking seriously rather than dismissing.

What is verified: the Vision-Exp release

The announcement page above is the entire official record: DeepSeek-V4-Flash-Vision-Exp is live on DeepSeek's API platform as of August 21, 2026, model ID deepseek-v4-flash-vision-exp, and nothing about weights, architecture, or a license is published anywhere on it. DeepSeek has not said the words "open source" about this model. The page is strictly an API announcement.

What the model does is verified and useful. It takes text plus images — JPEG, PNG, GIF, or WebP — via base64 inline, an external URL, or DeepSeek's new Files API, which is free and lets you upload once and reuse by file_id. It returns text. Its text capabilities match the text-only DeepSeek-V4-Flash: the same agent, reasoning, and world-knowledge behavior, on a 1M-token context with up to 384K tokens of output. Images are billed as tokens, up to 384 tokens per image, at exactly the V4-Flash list price. DeepSeek Harness 0.1.1 shipped the same day with native support.

The numbers DeepSeek reported

On multimodal agent benchmarks, the gain over the text-only model is the headline. Per DeepSeek's own figures — vendor-reported, unreproduced by anyone independent yet:

• ApexBench (Pass@1) — 36.5 for Vision-Exp vs 26.2 for text-only DeepSeek-V4-Flash.

• Agents' Last Exam — 27.3 vs 25.2.

• Chartography — 64.3.

• ZeroBench (Pass@5) — 35.0.

DeepSeek's summary line is that the model "brings multimodal agent performance close to Opus-4.8" while matching V4-Flash on pure-text benchmarks. Those are DeepSeek's figures with no independent evaluation yet, and "close to Claude Opus-4.8" is DeepSeek's framing, not a measured ranking.

The scoreboard above is the state of play as of August 22, 2026: the only released V4 vision model is the experimental, API-only one, and everything about the final model — weights, sizes, date — is expectation, not fact.

Why the open-weights question matters

DeepSeek is one of the few frontier labs whose production models routinely ship open, and that history is the real reason a leak about weights is news. The text-only DeepSeek-V4-Flash-0731 backbone is MIT-licensed — which is precisely why the community could build on it at all. If the final V4 vision model follows DeepSeek-VL2 and the text line, open weights change what a reader can do: local deployment on your own hardware, fine-tuning on proprietary image data, air-gapped use, and zero per-token cost on your own GPU.

There is also a reported roadmap that lines up. Chinese tech media reported in May 2026 that DeepSeek had said open-source V4 vision weights were planned for Q3 2026. That is a reported commitment, not a confirmed date — and the August 21 API-only step is consistent with "experimental first, weights later." On licensing, community coverage points to Apache-2.0 or MIT, but nothing is confirmed.

The community's interim answer

Because the official vision weights are not open, the community has built its own. Three notable projector-only grafts exist on the MIT-licensed DeepSeek-V4-Flash-0731 text backbone, each freezing the text model and training a small vision adapter in front of it:

• FlyCockpit/DeepSeek-V4-Flash-0731-vision — a DeepEncoderV2 vision tower plus a trained projector, served through vLLM via a small plugin; tested on NVIDIA DGX Spark clusters at roughly 50 tokens/sec.

• jarrelscy/deepseek-v4-flash-0731-vision — a frozen MoonViT tower (originally from Kimi-K2.6) plus a roughly 40M-parameter projector trained with GRPO, following the "WebBrain patchmerger" recipe.

• webbrain-one/DeepSeek-V4-Flash-0731-Vision-NVFP4 — an NVFP4-mixed-precision checkpoint with a routing-aware PatchMerger and custom SGLang integration.

These are unofficial, grafted, and not DeepSeek's model. They demonstrate that giving V4 Flash eyes is feasible, and they will be superseded the day official vision weights land. The fact that they exist at all is the clearest signal of how much the community wants those weights.

What V4-Vision-Final could look like

Everything in this section is speculation, because nothing is announced. The likeliest shape, from DeepSeek's own patterns: a final vision model — possibly at both Flash and Pro sizes — with open weights. The economics make local deployment plausible at the Flash size: DeepSeek-V4-Flash is a 284B-parameter mixture-of-experts model that activates only 13B per token, with a 1M-token context, and community deployment reports point to dual RTX 4090s or a single RTX 5090 for a quantized build. DeepSeek-V4-Pro is reported at roughly 1.6T total / 49B active — a bigger ask, but the same MoE logic applies. DeepSeek has also said V4 was trained multimodal from the start; Vision-Exp is the first surface of that capability, and the final model could expose more of it.

None of that is confirmed. It is the shape a reader should expect, not a promise.

How to play the wait

Here is what is true today, as distinct from what may be true later. DeepSeek-V4-Flash-Vision-Exp is live and routable right now, at DeepSeek's own list price — $0.15 per 1M input tokens and $0.29 per 1M output tokens, the same price as text-only V4-Flash, with images billed as tokens on top. OrcaRouter routes it through one API key across 200+ models, passes the provider list price through at 0% markup, and can fail over to a proven model automatically. That is the honest way to try an experimental model: point a test workload at Vision-Exp, keep a reliable fallback on the same endpoint, and let automatic failover swap in the dependable model the moment the experimental one stalls.

The model page above is where DeepSeek-V4-Flash-Vision-Exp is actually callable today — $0.15 in, $0.29 out, one API key, no second contract.

When the final open-weights model ships, two things happen at once: self-hosters get to run it locally, and the hosted route stays on the same key at the same pass-through price. You will not have to choose at announcement time — a routing layer means both options stay one call apart. To be clear about the boundary: OrcaRouter routes APIs; we do not host model weights. Local deployment of the final V4 vision model is about DeepSeek's weights, not about us.

What to watch next

Four things would turn this leak into a launch. An official naming and release date for the final V4 vision model. Confirmation of the license — MIT and Apache-2.0 are both plausible, and the difference matters for derivative work. A first independent benchmark of the final model, which would replace DeepSeek's "close to Claude Opus-4.8" with a number someone else verified. And, for the skeptical view, whether DeepSeek-V4-Flash-Vision-Exp itself ever gets weights — the leak says no, and DeepSeek's pattern supports that reading.

Until one of those lands, treat the leak as a directional signal rather than a fact. The experimental V4 vision model is real, useful, and API-only today. The open-weights final model is a well-sourced expectation, not a shipping product. If the watcher on X is right, the V4 vision line ends up where DeepSeek's models usually do — running on your own hardware, with the hosted API as the zero-ops alternative.