Product case study · September 1, 2026
DeepSeek V4 Flash Vision Exp is multimodal, open-weight, and not readiness-tested
DeepSeek calls this its first experimental multimodal V4 model. The MIT weights are public, but our nine-run practical benchmark remains unmeasured.

Treat DeepSeek V4 Flash Vision Exp as a research candidate, not a production recommendation. The original checkpoint remains unmeasured here; Hugging Face now lists a provider, while DeepSeek redirects the legacy API name to Flash-1017.
Sources checked October 27, 2026. This refresh checks public documentation and the local benchmark dossier; it does not report a model inference test. The September access block below is historical, and the nine runs remain unmeasured.
DeepSeek calls deepseek-ai/DeepSeek-V4-Flash-Vision-Exp the first experimental multimodal model in the DeepSeek-V4 series.[1] That wording is a publisher claim, not an independent finding. The repository is nevertheless concrete evidence that a multimodal checkpoint exists: Hugging Face labels its pipeline image-text-to-text and records the license as MIT.[1][2]
What the primary sources establish
The Hugging Face API record reports 304,646,824,126 parameters and 167,819,616,863 bytes of used storage.[2] At this page’s September 1, 2026 check it gave revision SHA 31ea11185e11ccafad1c385104188a9e3b648ad6, with creation dated August 31, 2026 and last modification dated September 1, 2026.[2] These are repository metadata, not claims about active parameters, memory consumption, inference speed, or minimum hardware.
The October 27 API metadata check still reports those parameter and storage totals, but now returns revision SHA 6821d6ad3681a4b137b066b76094fa82ebd0a380.[2] The dossier remains pinned to its original SHA; a newer repository revision is not automatically the same evaluation target.
The model card describes the release as experimental and multimodal.[1] We preserve that scope: this page does not infer a production support level from the pipeline tag, and it does not turn DeepSeek’s own examples or evaluations into independent benchmark results.
What remains unknown
At the original September 1 check, the dossier recorded no Hugging Face Inference Provider and no approved alternative route or hardware.[4] That provider statement is no longer current: the October 27 model page lists DeepInfra for image-text-to-text.[1] A listing does not verify account access, exact served revision, supported image formats, or a successful request. We have not tested that route and do not quote a price or measured latency for the historical checkpoint.
The repository size is also not a hardware prescription. Quantization, runtime support, memory overhead, parallelism, image preprocessing, and usable sequence length all affect a real deployment. Those variables need a documented serving setup and an actual run.
The old API name no longer selects this checkpoint
DeepSeek’s current documentation says requests using deepseek-v4-flash-vision-exp are served by DeepSeek-V4.1-Flash-1017, with Flash billing. The original hosted model is retired. The recommended current alias is deepseek-flash.[3]
This matters when testing a saved Hermes, Codex, or other agent configuration: keeping the old model string does not reproduce the original experiment. A response obtained through that alias today belongs in a Flash-1017 evaluation, not in this checkpoint’s nine-run dossier. See the Flash-1017 release analysis for the separate release.
| Route you are considering | What to establish before testing |
|---|---|
| DeepSeek API with the legacy Vision Exp name | Documentation maps it to Flash-1017; label the evaluation accordingly.[3] |
| Hugging Face’s listed DeepInfra provider | Confirm the provider’s actual checkpoint, revision policy, image support, and account access.[1] |
| Self-hosted original weights | Pin the checkpoint revision, runtime, image preprocessing, and supported settings. |
The model card now includes reference inference instructions and vLLM/SGLang examples. Its vLLM example uses four GB300 GPUs; this is a publisher example, not a minimum requirement or a Superbash-tested setup.[1] Repository size alone still cannot answer whether your machine can run it.
The reproducible benchmark is unmeasured
The dossier in benchmarks/deepseek-v4-flash-vision-exp/ fixes three multimodal tasks using original MIT-licensed repository fixtures: chart-to-structured-JSON extraction, screenshot-guided UI bug localization, and image-plus-repository change planning. The third task pairs a synthetic target screenshot with a small benign TSX source fixture and requires a plan grounded in both. Each task has one prompt and rubric and is scheduled for three independent runs under a pinned configuration.[4]
All nine raw result entries say not_run; raw output, timing, token, cost, environment, review, and result fields are null rather than zero.[4] The recorded September reason was access, not model quality: that evaluation had no approved compatible route or hardware. No weights were downloaded or inference attempted for that dossier. This refresh leaves those historical records unchanged; the new provider listing does not fill in missing results.[4]
The gate is deliberately resistant to one lucky output. A task passes only when at least two of its three runs are accepted against its rubric. A practical-readiness verdict requires at least two of the three tasks to pass. Because zero runs exist, the conclusion is no practical readiness verdict—blocked pending real runs.[4]
A practical preflight before spending on inference
These are proposed checks, not a successfully tested deployment recipe.
- Choose the evaluation target. Record either the historical checkpoint and revision or today’s hosted Flash version. Do not silently substitute one for the other.
- Validate the existing packet locally. From the repository root, run the command below. It checks the dossier, not model quality or endpoint access.
- Check input compatibility. The packet uses SVG fixtures. Confirm the route accepts them; if it requires raster images, record the rendering dimensions and conversion procedure as an explicit protocol change. Keep the TSX context with the third task.
- Check settings before a scored run. The protocol requests temperature 0, seed 20260901, 512 output tokens, no tools, and a fresh session for each run.[4] Record unsupported settings and any truncation; do not silently change the budget between attempts.
- Save evidence and score it. Preserve raw responses, elapsed time, usage, actual charges, and runtime details. The protocol calls for two independent reviewers and recorded adjudication.[4] Missing measurements stay null.
node benchmarks/deepseek-v4-flash-vision-exp/validate.mjs
If the provider cannot identify the served checkpoint or honor the protocol, keep the historical dossier unmeasured and design a separately labelled evaluation. If image input fails, fix and document the transport before interpreting an answer as evidence of visual understanding.
Can I reproduce Vision Exp by keeping its old API ID?
Not on DeepSeek’s current hosted API: the documented alias now routes to Flash-1017.[3] Configuration names alone do not establish model identity.
Does a listed provider mean the nine runs passed?
No. Availability and accepted task results are different evidence. The checked manifest still has nine not_run entries, with no measured readiness verdict.[4]
Use the lesson on evaluating open-weight models to turn these checks into a bounded comparison.
Sources
Checked October 27, 2026. Public documentation and repository records only; no authenticated inference request was made.
[1] https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-Vision-Exp — DeepSeek model card and Hugging Face deployment status
[2] https://huggingface.co/api/models/deepseek-ai/DeepSeek-V4-Flash-Vision-Exp — Hugging Face model API metadata
[3] https://api-docs.deepseek.com/quick_start/pricing/ — Current model mapping and legacy-name retirement
[4] https://github.com/rivermoonice/boxai4/tree/dbf69bc/benchmarks/deepseek-v4-flash-vision-exp — Superbash’s original protocol, fixtures, validator, and unmeasured run manifest; local records rechecked October 27, 2026
Put this to work
Keep publisher claims, repository metadata, and independently reproduced task results in separate evidence buckets.
Try
Verify the served checkpoint and image format before running the fixed fixture packet on a compatible route.
Prove it worked
Require two accepted runs out of three for a task, then two passing tasks out of three before making a readiness verdict.
Where it can pay
Do not estimate deployment economics until route, hardware, latency, usage, and correction time are measured.
Keep in view
- DeepSeek describes this as the first experimental multimodal DeepSeek-V4 model.
- Hugging Face reports 304,646,824,126 parameters, 167,819,616,863 bytes of used storage, an MIT license, and the image-text-to-text pipeline.
- Hugging Face now lists DeepInfra; the legacy DeepSeek API name serves Flash-1017, not the original checkpoint.