DeepSeek publishes V4 Flash Vision experimental weights under MIT

Quick answer

DeepSeek published the weights, tokenizer, prompt encoder, and minimal inference code for DeepSeek-V4-Flash-Vision-Exp on Hugging Face on August 31, 2026. The experimental model adds image understanding to the V4 Flash architecture and is licensed under MIT. Its configuration declares a 1,048,576-position context and a 32-layer vision encoder. The exact model was already available through DeepSeek's API from August 21, but local deployment now has an official weight repository. DeepSeek's benchmark comparisons are first-party results, and availability in Chat AI has not been verified.

Download Chat AI Opens the official App Store or Google Play for your device.

September 10 update: V4.1 Flash replaces the hosted experimental model

DeepSeek officially released V4.1 Flash on September 10, 2026, with native visual understanding. Its API now recommends the model name deepseek-flash. The older V4 Flash and V4 Flash Vision Exp hosted models are retired: requests using deepseek-v4-flash or deepseek-v4-flash-vision-exp are temporarily redirected to V4.1 Flash. Keeping the old name therefore does not preserve the old model's behavior. Self-hosting the older downloadable checkpoint remains a separate deployment path.

Sources: DeepSeek, DeepSeek

V4.1 Flash weights are also published

DeepSeek's official Hugging Face repository now lists V4.1 Flash weight shards, a technical report, tokenizer files, and inference and evaluation code. The public, ungated repository is marked with the MIT license. This is a separate checkpoint from the V4 Flash Vision Exp release described below: its architecture, configuration, and deployment requirements should be checked independently. This article has not run either checkpoint locally.

Sources: DeepSeek on Hugging Face

September 14 update: V4 Pro API service will continue

DeepSeek's updated official changelog says it will keep providing V4 Pro API service after September 14, 2026, with its billing method unchanged, in response to user demand. This supersedes the earlier plan described here to redirect deepseek-v4-pro requests to V4.1 Flash on September 14. DeepSeek says it will give further notice of changes. The separate retirement of V4 Flash and V4 Flash Vision Exp remains in the changelog; their legacy API names still temporarily route to V4.1 Flash. This update reflects provider documentation checked on September 14, not an independent test of live API routing.

Sources: DeepSeek

DeepSeek has released the experimental vision model's weights

The official Hugging Face repository contains the model weights, tokenizer, prompt-encoding reference, configuration, and minimal PyTorch inference code for DeepSeek-V4-Flash-Vision-Exp. DeepSeek describes it as the first experimental multimodal model in the V4 family. It builds on V4 Flash with visual modules and continued training for image understanding while retaining text and agent capabilities.

Sources: DeepSeek on Hugging Face

The model accepts images alongside text

DeepSeek's API documentation lists JPEG, PNG, GIF, and WebP inputs. Applications can send an image as base64 data, a public URL, or a file previously uploaded through the Files API. DeepSeek says the model can describe images, read screenshots, and analyze charts. Image inputs are converted to tokens based on their dimensions and billed together with text when the hosted API is used.

Sources: DeepSeek, DeepSeek

The configuration exposes a one-million-token context and a dedicated vision encoder

The published configuration sets max_position_embeddings to 1,048,576. It also describes a 32-layer vision encoder with a 1,024-dimensional hidden representation, 16 attention heads, and a maximum of 384 vision tokens. Those values document the checkpoint's configuration; they do not guarantee that every serving stack, request type, or hardware setup can use the full context efficiently.

Sources: DeepSeek on Hugging Face

The repository includes a local inference path, but deployment is not lightweight

DeepSeek's repository includes prompt encoders for OpenAI-style message blocks and compact image-path notation, plus a minimal inference implementation covering the vision encoder, aligner, mixture-of-experts backbone, DFlash attention, Hyper-Connections, and DSpark forward path. The Hugging Face repository is roughly 168 GB and contains dozens of weight shards, so a developer should check storage, accelerator memory, runtime compatibility, and supported quantization before treating this as a local-laptop model.

Sources: DeepSeek on Hugging Face

The weights and reference code use the MIT license

The repository is published under the MIT license, which permits use, modification, distribution, sublicensing, and sale subject to preserving the required notice and license terms. That broad software license does not remove obligations that may apply to input data, generated output, privacy, safety, or downstream product use.

Sources: DeepSeek on Hugging Face

DeepSeek's benchmark table is useful context, not independent verification

DeepSeek reports that the vision checkpoint improves on V4 Flash in its selected multimodal agent evaluations while remaining comparable on text-only agent tasks. The model card documents the evaluation settings and comparison table, but these are provider-run results. Teams should test image reading, chart interpretation, tool use, accuracy, latency, and operating cost on their own workloads before choosing the model.

Sources: DeepSeek on Hugging Face, DeepSeek

Availability in Chat AI has not been verified

DeepSeek-V4-Flash-Vision-Exp is a distinct experimental checkpoint. The exact model does not appear in Chat AI's currently verified model directory, so this article does not claim that it is available in Chat AI. Model access can also change across the DeepSeek API, self-hosted deployments, and third-party platforms.

Sources: Chat AI, DeepSeek on Hugging Face

Frequently asked questions

What readers usually ask

Is DeepSeek V4 Pro being retired on September 14?

DeepSeek's updated official changelog says V4 Pro API service will continue after September 14, 2026, with its billing method unchanged. That replaces the earlier announced routing plan. The change does not restore the retired V4 Flash or V4 Flash Vision Exp hosted models.

Does the old Vision Exp API name still select the experimental checkpoint?

No. DeepSeek's September 10 release notice says deepseek-v4-flash-vision-exp temporarily routes to V4.1 Flash. Use deepseek-flash for the current hosted model; the older downloadable checkpoint is distinct from the hosted API.

What is DeepSeek-V4-Flash-Vision-Exp?

It is DeepSeek's first experimental multimodal model in the V4 family. It combines the V4 Flash language-model architecture with visual modules so it can process images alongside text.

Are the DeepSeek V4 Flash Vision weights available to download?

Yes. DeepSeek has published model weights, configuration, tokenizer files, prompt encoding, and minimal inference code in its official Hugging Face repository.

What license does the model use?

The official repository uses the MIT license. Developers should read the license and separately review obligations for data, privacy, safety, and downstream use.

Which image formats does the DeepSeek API accept?

DeepSeek documents JPEG, PNG, GIF, and WebP. Images can be supplied as base64 data, a public URL, or a Files API reference.

Does the model support a one-million-token context?

Its published configuration declares 1,048,576 maximum positions. Effective usable context and performance can still depend on the serving stack, hardware, request shape, and implementation.

Is DeepSeek-V4-Flash-Vision-Exp available in Chat AI?

Availability has not been verified. The exact experimental model is not listed in Chat AI's current verified model directory.

Evidence