DeepSeek-V4-Flash-Vision-Exp

On August 21, 2026, DeepSeek released an experimental multimodal model called deepseek-v4-flash-vision-exp on its API Platform. Built on DeepSeek-V4-Flash, it keeps the same text performance (agentic tasks, reasoning, world knowledge) while adding image understanding. There is no price premium — it is billed at the same rate as V4-Flash. Despite the "Exp" (experimental) label signaling that DeepSeek itself does not treat it as production-guaranteed, it is available in the API immediately.

Ten days after the API launch, DeepSeek published the deepseek-v4-flash-vision-exp weights on Hugging Face on August 31, 2026, under a plain MIT license. It's the first open-weight multimodal release in the V4 family, with no restrictions on commercial use, modification, or redistribution.

The checkpoint builds on the DeepSeek-V4-Flash backbone (284B total parameters, 13B activated per token, 43 all-MoE layers with 256 routed experts plus one shared expert, top-6 routing, and a native 1M-token context window), adding a 32-layer vision encoder and a two-layer aligner on top. The full published checkpoint runs about 304.6B parameters, and each image still compresses to a maximum of 384 tokens — consistent with how the API bills image inputs.

To check whether it fits on hardware you actually own, the VRAM calculator (free, no signup) lets you pick a model, quantization level, and context length and get a real memory estimate. For the underlying math on context length and memory, see the KV cache and context length VRAM guide.

- The official ~168GB FP4/FP8 checkpoint does not run on a single GPU. It needs either multiple GPUs or CPU offload with a large amount of RAM
- Community GGUF quantizations (from projects like Unsloth) shrink it to as little as 82.4GB, but that still doesn't fit on a typical consumer GPU
- Realistically your options are: a multi-GPU workstation or server, a CPU/GPU offload setup with ample RAM, or waiting for a more aggressive quantization to appear. For a sense of what a comparable-scale local build looks like, see the Solar Open2 250B requirements

Inference engine support is still catching up too — image input needs a vision-enabled llama.cpp build or equivalent. For the broader landscape of local inference engines, see the local inference engine comparison.

MIT effectively allows unrestricted modification, fine-tuning, redistribution, and commercial hosting of the weights. Still, running a 168GB-class checkpoint yourself costs more than the API for most small and mid-sized teams. The practical decision comes down to two questions: if the API already meets your latency, cost, and data-residency needs, there's no reason to switch. If you have a hard requirement to keep data in-house and already have (or can invest in) multi-GPU infrastructure, that's when going after the open weights pays off. For a middle ground that mixes cloud API and local inference, see hybrid cloud/local LLM cost reduction.

Adding image understanding opens up multimodal agent tasks that text-only V4-Flash could not handle: reading a screenshot to plan a UI action, summarizing numbers from a chart or dashboard image, or extracting fields from a scanned invoice or document. DeepSeek states it made a large jump from V4-Flash on multimodal agent benchmarks.

- UI-operating agents: parsing screenshots and planning the next action
- Document processing: extracting fields from invoice or form images
- Chart reading: summarizing numbers from graphs or dashboard screenshots
- Mixed text-and-image chat: combining text and images within a single conversation

The key pricing point is that there is no premium for image input. deepseek-v4-flash-vision-exp bills at the same rate as V4-Flash, and images are counted at the same per-token price as text. The catch is that each image is tokenized at up to 384 tokens, so image-heavy requests will add meaningfully to input token counts.

- One image = up to 384 tokens counted toward billing
- Image tokens are simply added to the text prompt's token count
- The per-token rate itself matches V4-Flash's text rate — no separate image surcharge
- Uploading to the Files API itself is free (only tokens consumed at inference time are billed)

The fastest path is sending a message that mixes text and images to the Chat Completions (or Messages / Responses) endpoint. Images can be passed as base64, an external URL, or via the Files API. Below is a minimal curl example using Chat Completions with an externally hosted image.

There are three ways to pass images — base64, external URL, and the Files API — each suited to different workflows. For workflows that reference the same image repeatedly, the Files API is more bandwidth-efficient since you upload once and reuse the resulting file_id.

A practical detail worth noting is that the Files API is now available for deepseek-v4-flash-vision-exp. Upload an image once to get a file_id, then reference that file_id in subsequent requests instead of re-sending base64-encoded image data every time. This saves bandwidth for batch jobs or agent loops that repeatedly touch the same image, and using the Files API itself carries no charge.

DeepSeek describes the model as approaching Claude Opus 4.8 on multimodal agent benchmarks. This is a vendor-reported figure that has not been independently reproduced or verified, so it should be treated with appropriate caution. Moreover, according to DeepSeek's own published table, it only surpassed Opus 4.8 on 3 of 11 benchmarks — it does not outperform Opus 4.8 across the board. "Approaching" should be read as strong performance on select metrics, not overall superiority.

- This is a vendor-reported figure; independent reproduction has not been confirmed as of this writing
- It surpassed Opus 4.8 on only 3 of 11 benchmarks, matching or trailing on the remaining 8
- "Approaching" refers to select metrics, not across-the-board superiority
- On text-only tasks, performance is described as equivalent to the base V4-Flash

When comparing options internally, text-centric workloads are still well served by DeepSeek-V4-Flash, and DeepSeek-V4-Pro is the candidate for heavier tasks that need higher accuracy — V4-Pro's own GA release and pricing update landed on August 13, 2026. Vision-Exp becomes worth considering only once image understanding is actually needed, and for teams wanting still higher accuracy, closed models like Claude Opus 4.8 come into the comparison. It's also worth reviewing the hardware and VRAM requirements across the V4 lineup.

- Spec-change risk: as an "Exp" (experimental) model, behavior and specs may change without notice
- No self-hosting: weight availability for Vision-Exp has not been confirmed as of this writing, so on-premises deployment is not an option
- Availability and rate limits unconfirmed: whether it matches standard V4-Flash availability guarantees is not publicly stated
- Image token cost management: for image-heavy workflows, estimate the up-to-384-token-per-image cost in advance

Given its experimental status, it's safer to avoid wiring it directly into a primary production path. Instead, run it alongside existing flows on the subset of workloads that actually need image understanding, confirm accuracy and cost in practice, and only then widen the rollout.