DeepSeek ships V4-Flash-Vision-Exp — its first multimodal model, matching V4-Flash on text, adding image input

On 2026-08-21, DeepSeek launched deepseek-v4-flash-vision-exp — its first multimodal model — on the DeepSeek API platform (news post, 2026-08-21). The model matches V4-Flash on text capabilities (agents, reasoning, world knowledge) and adds image input at the same per-token pricing. A new Files API lets you upload images once and reuse them across requests.

The same day, DeepSeek Harness 0.1.1 shipped with out-of-the-box support for the new model (Harness quickstart, 2026-08-22).

What it is

Experimental, not production. The -Exp suffix is deliberate — DeepSeek flags this as an experimental release. It supports Thinking and non-thinking modes, 1M context length, JSON output, tool calls, the Responses API, and the Anthropic API (pricing page, 2026-08-22).

Image input via three methods:

• Base64 inline — encode and embed directly in the request (48 MiB body limit)
• External URL — pass an http(s) link; max 8192 characters, 32 MiB file, 60-second download
• Files API file_id — upload once, reference by ID; supports up to 64 MiB per image

Supported formats: JPEG, PNG, GIF, WebP. Up to 600 images per request. Images in user messages only — system and assistant messages return a 400 error (vision guide, 2026-08-22).

Token billing for images. Every image is resized before inference. Images below ~384x384 pixels are scaled up; larger images are scaled down to roughly 800x800 pixels. The upper bound is 384 tokens per image, billed as input tokens at V4-Flash rates (vision guide, 2026-08-22).

Pricing (verified 2026-08-22)

V4-Flash-Vision-Exp is priced identically to V4-Flash (pricing page, 2026-08-22):

At 384 tokens per image (worst case), the image cost per request is roughly $0.00008 off-peak — negligible compared to text tokens on most prompts.

Files API: upload once, reuse everywhere

The Files API launched alongside the vision model. It is free to use (Files API guide, 2026-08-22):

• Upload an image once, get a file_id (format: file-api-...)
• Reference that ID in any subsequent chat request — no re-upload needed
• Max upload size: 64 MiB. Max stored files: 10,000 per user. Max storage: 25 GiB
• Optional expiration: 1 hour to 30 days, or permanent (default)
• Supported formats: JPEG, PNG, GIF, WebP

The API works with both the OpenAI-compatible endpoint and the Anthropic-compatible endpoint (requires anthropic-beta: files-api-2025-04-14 header).

Benchmark claims: close to Opus-4.8 (vendor-reported)

DeepSeek’s news post claims V4-Flash-Vision-Exp “brings multimodal agent performance close to Opus-4.8” on multimodal agent benchmarks (news post, 2026-08-21). The post includes a benchmark chart (/img/v4_260821_benchmark_en.png) but does not publish the underlying numbers in text.

This is a vendor-reported claim — no independent reproduction exists. The article does not assert the model matches or beats Opus-4.8; it quotes DeepSeek’s characterization. No primary-source head-to-head comparisons vs GPT-4V, Claude, or Gemini were published alongside this release.

What changed since V4-Flash

• Image input added — V4-Flash was text-only; V4-Flash-Vision-Exp adds multimodal understanding
• Pricing unchanged — same per-token rates as V4-Flash
• Files API launched — new infrastructure for image upload and reuse across requests
• Harness 0.1.1 — same-day release with out-of-the-box model support

Prior DeepSeek coverage: dsh agent harness (2026-08-19).

Risks and caveats

• Experimental. The -Exp suffix means this is not a production-stable API. Expect changes.
• Benchmark claim is vendor-reported. “Close to Opus-4.8” comes from DeepSeek’s own evaluation. No independent third-party benchmarks exist yet.
• No competitor comparisons published. The news post does not include head-to-head data vs GPT-4V, Claude, or Gemini.
• Image token billing adds up. At 384 tokens per image, a 10-image prompt costs ~3,840 input tokens — about $0.0008 off-peak. Multiply by request volume.
• Files API is free but has limits. 10,000 files, 25 GiB storage, optional 30-day expiration. Plan for cleanup if uploading at scale.

What to watch

1. Stable release timeline — when does -Exp become GA?
2. Independent benchmarks — third-party evaluations on multimodal agent tasks
3. Pricing stability — will V4-Flash-Vision-Exp stay at V4-Flash rates after experimental phase?
4. Harness adoption — does 0.1.1 support drive real usage of the vision model?
5. Competitor responses — how do GPT-4V, Claude, and Gemini pricing/capabilities shift?

Sources