DeepSeek published open weights for V4-Flash-Vision-Exp on August 31, 2026, closing a gap its own API changelog had tracked since April: the V4 family could reason over text and code, but it could not look at an image. The release makes a 305-billion-parameter vision-capable sparse Mixture-of-Experts model freely downloadable under an MIT license — the first model in the class that a developer can self-host, fine-tune, and deploy outside China without paying a license fee. Whether those weights are worth the multi-GPU infrastructure they require depends on benchmark figures from its own harness, and that no independent laboratory has reproduced as of publication.

The release arrives one week after a competing open-weight multimodal model in the same size class — Z.ai’s GLM-5.3-Flash, published on Hugging Face with its weights on August 25 — marking the first week in which two Chinese labs simultaneously offered MIT-licensed multimodal MoE checkpoints above 300 billion parameters. DeepSeek’s choice to run a ten-day API-only window before publishing the weights differs from its July 31 release of the text-only V4-Flash-0731, whose V4-Flash-0731 same-day weights appeared alongside the API update.

What Was Released — and What Changed Inside the Weights

V4-Flash-Vision-Exp keeps the text backbone of V4-Flash-0731 unchanged: a 284-billion-parameter sparse MoE transformer with roughly 13 billion parameters active per forward pass, 43 layers, a hidden size of 4,096, and the same one-million-token context window that the text model made standard for the V4 family. The model card confirms expert routing selects 6 of 256 routed experts per token, plus one shared expert that is always active.

The vision capability arrives through two added modules. A 32-layer, 1,024-dimensional encoder with 16 attention heads processes images by breaking them into 14-by-14-pixel patches. The encoder feeds an aligner that projects visual features into the language backbone’s 4,096-dimensional hidden space, where they are treated as a sequence of tokens alongside the text input. The 384-token image budget analysis confirmed — the same figure that appears in the API billing specification, where images are charged at up to 384 tokens each at standard V4-Flash rates. The architectural ceiling and the billing ceiling are identical because one determined the other: the pricing policy is a direct expression of how many tokens the encoder can generate.

The two added modules bring total parameter count to approximately 305 billion, as reported by the Hugging Face repository metadata — a difference of roughly 21 billion parameters above the 284-billion MoE backbone.

A meaningful precision difference from the text-only model. V4-Flash-0731 ships with its MoE expert weights in FP4 precision and other parameters in FP8 — a mixed-precision scheme that FP4 checkpoint fits 170-175 GB and requires NVIDIA Blackwell-generation B200 GPUs for the FP4 expert kernels. V4-Flash-Vision-Exp ships entirely in FP8 (e4m3 format, 128-by-128 blocks with ue8m0 scale), with no FP4 expert weights — confirmed by the Hugging Face model card, which lists tensor types as BF16, F32, F8_E4M3, I8, and I64, with no FP4. The vision model therefore cannot run the FP4 MoE expert kernels available to the text model, runs on both Blackwell and Hopper hardware through the FP8 backend, and requires significantly more VRAM than the text-only Flash baseline. The vision model community GGUF conversions start at approximately 155 GB at 4-bit quantization — that is a multi-GPU or high-memory workstation decision, not a single consumer GPU.

What the Benchmark Table Actually Shows

DeepSeek’s comparison against V4-Flash-0731 and Opus-4.8 was run using Harness Minimal Mode, the lab’s own agent evaluation framework, at max reasoning effort with temperature 1.0 and top_p 0.95. The harness code has not been published, which means the figures cannot currently be reproduced by an independent laboratory.

The multimodal benchmark gains require careful reading because DeepSeek flags the comparison in a footnote: on ApexBench and Agents’ Last Exam, the text-only V4-Flash-0731 baseline “ignores the multimodal elements in the input.” Part of the advertised leap on those benchmarks is therefore structural — a text model that skips image inputs will underperform on tests built around visual reasoning regardless of its text capability. The more meaningful comparison is against Opus-4.8 on tasks where both models process identical input.

Against Opus-4.8, V4-Flash-Vision-Exp wins three of eleven benchmarks tested: DeepSWE at 59.3 versus 58.0, Agents’ Last Exam at 27.3 versus 25.7, and ZeroBench Pass@5 at 35.0 versus 34.0. It stays within roughly one point on Terminal Bench 2.1 (83.9 versus 85.0), Toolathlon-Verified (75.9 versus 76.2), and Chartography (64.3 versus 65.0). The gaps widen on the two hardest long-horizon tasks — NL2Repo (57.7 versus 69.7) and DSBench-Hard (63.6 versus 71.7) — indicating that tasks requiring synthesis across many steps and documents still favor the heavier Opus-4.8 model. Cybergym is the sole regression against the text-only baseline, at 75.3 against V4-Flash-0731’s 76.7.

The understated finding is on text. DeepSeek claims only that the vision variant “maintains comparable performance on text-only agent tasks,” but the official benchmark table shows V4-Flash-Vision-Exp ahead of V4-Flash-0731 on six of seven text benchmarks, with gains of 5.6 points on Toolathlon-Verified and 4.9 points on DeepSWE. Adding the vision encoder did not cost the text backbone anything measurable.

The precedent for treating self-reported DeepSeek benchmarks with caution is established: in independent testing by yage.ai, V4-Pro’s vendor-reported SWE-bench Verified score of 80.6 percent dropped to 8 percent pass@1 on the more rigorous DeepSWE benchmark with a 0.3 percent false-positive verifier, versus 8.5 percent for the vendor-controlled harness. The V4-Flash-Vision-Exp DeepSWE figure of 59.3 uses the tighter benchmark, which is meaningful — but it remains self-reported until a named independent laboratory publishes results from a separately configured evaluation.

Why the Ten-Day Window Matters

The release sequence distinguishes this checkpoint from V4-Flash-0731, whose weights appeared on Hugging Face the same morning as the API update in July. V4-Flash-Vision-Exp ran API-only from August 21 through August 30, then published weights on August 31 — a ten-day window during which DeepSeek could collect production telemetry on real multimodal workloads before releasing the checkpoint permanently. The API launched on August 21, 2026 to OpenRouter and Vercel’s AI Gateway the same day.

The distinction from the competing Z.ai release is worth stating precisely. GLM-5.3-Flash — a 320-billion-parameter multimodal MoE from Z.ai — published its open weights on August 25, six days before V4-Flash-Vision-Exp and four days after DeepSeek’s API launch. Both models arrived as MIT-licensed multimodal MoE checkpoints in the 300-billion-plus class within the same week, by different release strategies. The difference is sequencing, not commitment: DeepSeek chose a brief telemetry window; Z.ai published weights on day one.

The MIT license applied to V4-Flash-Vision-Exp is materially different from more restrictive weight licenses that other large model labs have used. It permits commercial deployment, fine-tuning on proprietary datasets with private retention of results, architectural modification, and redistribution of derivative works, with only a copyright notice required. For teams that need to adapt the vision encoder for domain-specific applications — medical imaging, satellite imagery, technical diagram parsing — the MIT license removes the legal friction that would otherwise require negotiating a commercial agreement.

How the “Exp” Suffix Should Be Read

DeepSeek has used the “Exp” designation consistently across the V4 family to mark checkpoints that are API-first, under active evaluation, and not yet in their final production form. V4-Flash-0731 released as the production-stable Flash checkpoint; the current vision checkpoint carries the “Exp” suffix. Whether a production-stable V4 vision model follows, and what that release looks like, has not been announced.

The benchmark evaluation set DeepSeek chose is also narrower than the broad visual question-answering suites that most multimodal model releases benchmark against. DeepSeek measured the model on agentic and chart-reading tasks — workloads relevant to software engineering and data analysis — rather than the general visual perception, OCR, and science reasoning benchmarks that establish a model’s position in the wider multimodal landscape. That choice makes the results directly useful for teams building agentic software pipelines and less informative for teams needing general visual reasoning capability.

What the China Data Law Means for This Release

DeepSeek is incorporated in Hangzhou, China, and its legal obligations under Chinese law apply whether a user accesses the model through the hosted API or through a self-hosted deployment.

For users of the hosted API at api.deepseek.com, prompts and conversation history transit to DeepSeek’s servers in China. China’s National Intelligence Law (2017), Article 7 mandates intelligence cooperation from all organizations and citizens — an obligation that applies without exception, without a transparency requirement, and without a mechanism for legal refusal. China’s Cybersecurity Law (2017) adds the authority to require technical assistance, and the Data Security Law (2021) adds cross-border data transfer restrictions and additional government-access provisions. These are fixed statutory requirements of DeepSeek’s operating jurisdiction, not interpretive risks to be weighed against price.

For users of the MIT-licensed open weights on self-hosted, non-Chinese infrastructure, the data routing concern is eliminated — prompts do not leave the operator’s own environment. Self-hosting does not, however, change DeepSeek’s own legal obligations under Chinese law, and no named independent security audit of the V4-Flash-Vision-Exp weights has been published. For teams evaluating whether to self-host as a data governance strategy, the self-hosting path removes one category of risk while leaving unresolved questions about what behavioral constraints, if any, are embedded in the weights themselves.

For regulated industries — healthcare, financial services, government-adjacent work — the hosted API is incompatible with most data governance frameworks in US jurisdictions. Multiple US federal agencies, including Navy and NASA banned DeepSeek on official systems. Teams without regulated-data obligations who deploy the open weights on their own infrastructure represent a structurally different use case.

Access and Hardware Reality

The weights are published at the deepseek-ai organization on Hugging Face, under the collection for DeepSeek-V4-Flash-Vision-Exp. The model accepts JPEG, PNG, GIF, and WebP image formats, with inline images up to 32 megabytes (MB) and Files API images up to 64 MB.

The API endpoint speaks both OpenAI Chat Completions and Anthropic Messages formats, which means teams already using either SDK can route to V4-Flash-Vision-Exp without a proxy layer. The API model identifier is deepseek-v4-flash-vision-exp.

Self-hosting requires meaningful hardware. The model card provides an official 4×GB300 vLLM deployment example with DSpark speculative decoding enabled. Unlike the text-only V4-Flash-0731, which achieves its compact 170-to-175-GB footprint through FP4 MoE expert weights on Blackwell GPUs, the Vision-Exp checkpoint ships entirely in FP8 — the same precision as the FP8-only quantized build of the text model, which occupies approximately 284 GB before the vision tower’s approximately 42-GB addition. Community 4-bit GGUF quantization options starting at approximately 155 GB are already appearing on Hugging Face, but running the model at near-native quality requires multi-GPU configurations that make this a team infrastructure decision rather than an individual setup.

For self-hosted deployments, vLLM requires --trust-remote-code until the custom attention modules land in upstream Transformers. SGLang supports DSpark speculative decoding via --speculative-algorithm DSPARK without a separate draft model path, since the draft and target weights share the same checkpoint.

No reasoning benchmark results using a third-party-configured harness have been published as of publication.

Frequently Asked Questions

How does V4-Flash-Vision-Exp compare to Opus-4.8 on image tasks?

On the benchmarks DeepSeek measured, V4-Flash-Vision-Exp leads Opus-4.8 on three of eleven tasks (DeepSWE, Agents’ Last Exam, and ZeroBench Pass@5) and trails it on long-horizon multi-step tasks such as NL2Repo and DSBench-Hard by 12 and 8 points respectively. All figures are self-reported using DeepSeek’s own evaluation harness. The benchmark set is narrow — focused on agentic software and chart-reading tasks — rather than the broad visual perception and science reasoning suites that the wider multimodal research community uses to position models. Independent reproduction of any of these figures has not yet been published.

What does it mean that V4-Flash-Vision-Exp ships in FP8 instead of FP4 like the text model?

The text-only V4-Flash-0731 uses FP4 precision for its MoE expert weights, which shrinks the native checkpoint to roughly 170 to 175 GB and FP4 weights need Blackwell B200 GPUs for the FP4 expert kernels. The vision model ships entirely in FP8 (no FP4), which means it is compatible with both Blackwell and Hopper H100 GPUs but requires substantially more VRAM — in the range of 326 GB in full FP8, or approximately 155 GB at community 4-bit quantization. Teams comparing the vision model’s self-hosting cost against the text model’s should use the FP8 quantized text baseline (approximately 284 GB) as the more accurate comparison, not the FP4+FP8 native build.

Is it safe to send sensitive data to the DeepSeek Vision API?

For regulated or sensitive data, it is not. Prompts sent to DeepSeek’s hosted API — including any images attached to requests — transit to servers in China. China’s Article 7 requires intelligence cooperation from all organizations without exception or transparency. That legal condition is fixed and cannot be addressed by DeepSeek’s privacy policy or by selecting a particular tier of the API. Self-hosting the MIT-licensed open weights on infrastructure outside China removes the data-transit concern, though it does not eliminate questions about training-time behavioral constraints in the weights themselves.

Will there be a production-stable V4 vision model after this experimental release?

DeepSeek has not announced one. The “Exp” suffix in the model’s name has consistently marked checkpoints that are under active evaluation and not yet in final form. Whether a production-stable V4 vision checkpoint will follow — and whether it will ship with FP4 MoE expert weights like the text-only model — has not been disclosed.