DeepSeek Opens V4-Flash-Vision-Exp Weights Under MIT

Introduction

DeepSeek published the open weights of DeepSeek-V4-Flash-Vision-Exp to Hugging Face on August 31, 2026. The repository's initial commit is timestamped 2026-08-31T06:16:18Z, and the LICENSE file in the repository is the MIT License.

The model is not new to users, only to self-hosters. DeepSeek's API changelog dates the hosted version to August 21, 2026, when it became callable as deepseek-v4-flash-vision-exp. What arrived on August 31 is the downloadable checkpoint plus a reference implementation, roughly ten days after API availability. The model card describes it as "our first experimental multimodal model in the DeepSeek-V4 family," built by "incorporating visual modules" into the DeepSeek-V4-Flash architecture and continuing training "to unlock visual understanding capabilities."

Feature Overview

A vision tower added to an existing text backbone. The config.json exposes a separate block of vision fields: 32 vision layers, a 1024-dimensional vision hidden size, 16 vision attention heads, a 2816 intermediate size, patch size 14, a downsample ratio of 3, a cap of 384 vision tokens per image, a 147,456-pixel minimum, and a maximum width-to-height ratio of 8.

The size of that addition is measurable. Hugging Face's safetensors index reports about 304.65 billion tensor elements for this checkpoint against about 304.18 billion for DeepSeek-V4-Flash-0731, and effectively all of the difference sits in the BF16 bucket: roughly 466 million elements. That is consistent with the model card's framing, a vision encoder and aligner bolted onto an otherwise unchanged backbone, rather than a retrained model. DeepSeek does not state a total or active parameter count for the vision variant in either the model card or the changelog, and the packed low-precision storage means the raw element count is not a clean parameter figure.

The MoE backbone is carried over unchanged. 43 hidden layers, hidden size 4096, 256 routed experts plus one shared expert, six experts routed per token, MoE intermediate size 2048, and a 129,280-token vocabulary. Weights ship in fp8 (e4m3 format, 128x128 block scaling) with expert_dtype set to fp4. The repository's files total roughly 168 GB.

Million-token context. max_position_embeddings is 1,048,576, reached via YaRN rope scaling with a factor of 16 over an original 65,536-token window. DeepSeek's pricing page lists a 1M context length and a 384K maximum output for all three V4 models.

A reference implementation, explicitly not a serving stack. The repo ships an encoding/ directory for prompt formatting and an inference/ directory whose README calls it "a readable reference implementation rather than a production serving engine." It covers "Vision + Aligner, DFlash, MoE, Hyper-Connections, and Transformer.forward_spec() for the DSpark forward path." A convert.py script converts Hugging Face weights into one checkpoint file per tensor-parallel rank, invoked with --n-experts 256 and --expert-dtype fp4; the worked example runs at MP=4. Two sample files express the same interleaved two-image prompt in an OpenAI-style JSON form and a compact <image>path</image> text form, and are documented as producing identical token IDs.

Priced the same as the text Flash model. On DeepSeek's API, the vision model carries identical rates to deepseek-v4-flash: $0.44 per million cache-miss input tokens and $1.32 per million output tokens at peak, halved off-peak to $0.22 and $0.66, with cache hits at $0.014 peak and $0.007 off-peak. Peak hours are 01:00-04:00 and 06:00-10:00 UTC, Monday through Friday. The concurrency limit is 2500. Images are converted to tokens based on their dimensions and billed as input. One capability gap is listed: FIM completion is marked "not supported" for this model, while both text V4 models support it in non-thinking mode.

DeepSeek-reported benchmarks. The model card publishes the following, with DeepSeek models evaluated under the minimal mode of DeepSeek Harness at max reasoning effort, temperature 1.0, top_p 0.95:

Usability Analysis

The multimodal gaps deserve a footnote that DeepSeek supplies itself. On ApexBench and Agents' Last Exam, the card marks the V4-Flash-0731 column with a dagger and explains that the text model "ignores the multimodal elements in the input." So 36.5 against 26.2 is not a like-for-like improvement; it compares a model that reads the images with one that cannot. The honest reading of the multimodal rows is that they establish a baseline, not a delta.

The text-side rows are the more informative comparison, and they are mixed. Toolathlon-Verified moves from 70.3 to 75.9 and DeepSWE from 54.4 to 59.3, while Cybergym slips from 76.7 to 75.3. That pattern matches the card's own claim of "comparable performance on text-only agent tasks" better than it supports a general upgrade, and it means the vision variant is an addition to the lineup rather than a replacement for the text Flash model.

Running it locally is a serious undertaking. Roughly 168 GB of weights, already in fp8 with fp4 experts, is multi-GPU territory before any context is allocated, and the shipped tooling assumes tensor parallelism rather than single-device inference. For most teams the practical entry point is the API, where the model has been live since August 21 at the same price as the text model. The weights matter for anyone who needs on-premise deployment, wants to study the vision-aligner design, or intends to build derivative work on an MIT-licensed multimodal MoE.

Pros and Cons

The release is unusually well documented for a checkpoint drop: a runnable reference implementation, a documented prompt-encoding path with equivalence tests, and a permissive MIT license with no gating agreement. The vision addition is small and legible, which makes the architecture easy to study.

The counterweights are real. The "Exp" in the name is DeepSeek's own signal that this is experimental, the benchmark table is entirely vendor-reported, and the two evaluation YAML files shipped in the repository under .eval_results/ cite the model card itself as their source, so they are not independent confirmation. No technical report is linked from the model card.

Outlook

DeepSeek's changelog documents a preview-to-GA path for this family already: V4-Flash entered public beta on July 31 and V4-Pro moved to GA on August 13 after a preview stage. An experimental vision checkpoint priced at parity with the text model, shipped both hosted and open, reads like the same pattern at an earlier stage. Whether the vision capability folds into the mainline Flash and Pro models or stays a separate branch is the open question the release does not answer.

Conclusion

A competent, well-packaged first multimodal entry in the V4 line, released open under MIT with enough reference code to actually reproduce the inference path. It is best read as DeepSeek establishing a multimodal baseline rather than claiming a lead: the text scores move a little in both directions, and the multimodal comparisons are against a model that was not looking at the images. Worth attention from teams that need self-hosted vision-language capability or want to study the aligner design; teams that only need the capability should use the API, where it has been available since August 21.

Editor's Verdict

DeepSeek Opens V4-Flash-Vision-Exp Weights Under MIT brings real, demonstrable value, though with caveats that deserve weighing.

The strongest case for paying attention: MIT license with no gating agreement, covering both the weights and the shipped reference code. That alone raises the bar for what readers should expect in this space. Reinforcing that, unusually complete packaging for a checkpoint drop: prompt-encoding module with equivalence tests, weight conversion script, and a readable inference path covering vision, aligner, MoE and the DSpark forward — practical value rather than just headline appeal. The broader signal worth registering is straightforward: the vision addition is measurable and small: Hugging Face reports about 304.65 billion tensor elements against 304.18 billion for DeepSeek-V4-Flash-0731, and effectively all of the ~466 million difference sits in the BF16 bucket, matching an encoder-and-aligner graft rather than a retrained backbone. On the other side of the ledger, one constraint is real rather than a marketing footnote: every published benchmark is vendor-reported on DeepSeek's own harness, and the evaluation files bundled in the repository cite the model card as their source rather than an independent run. It should factor into any serious decision. Layered on top of that, Cybergym regresses against the text V4-Flash-0731, from 76.7 to 75.3 — which narrows the set of teams for whom this is an obvious yes.

For multi-model deployment teams, cost-conscious operators, and developers willing to evaluate beyond the major labs, a measured trial makes sense, with clear criteria for when to expand or pull back. For everyone else, the safer posture is to monitor coverage and revisit once the use cases that matter to your team are demonstrated in the wild.