Skip to main content

DeepSeek V4-Flash-Vision: First Native-Vision Frontier Open Weight

Intermediate

At 00:00 UTC on August 31, 2026, DeepSeek quietly pushed a new repository to Hugging Face: deepseek-ai/DeepSeek-V4-Flash-Vision-Exp. Ten days after the same model went live on their API (August 21), the weights are now MIT-licensed and public. This is the first V4-family model with native vision — not a text model with a bolt-on adapter, but a joint-trained multimodal MoE — and its agent scores on the vision side jumped from 26.2 to 36.5 on ApexBench without moving the text-only agent scores at all. On several agent benchmarks it now matches or edges past Claude Opus 4.8 while the API costs $0.22 per million input tokens.

This page is the practical field guide the launch didn't come with: what changed architecturally, the numbers that matter, the one hard limit nobody warns you about, how to run it, and when Claude is still the right call.

What you'll learn
  • Understand what 'native vision' means here — the V4-Flash backbone plus a vision encoder + aligner trained together, not stapled on after
  • Read the benchmark table the way the release intends: text-only agents preserved, vision agents up sharply, both within striking distance of Claude Opus 4.8
  • Know the hard 384-tokens-per-image architectural cap and why it changes how you prompt
  • Do the honest cost math: $0.22 / $0.66 / $0.007 (cached) per million tokens vs Claude Opus 4.8 at $15 / $75
  • Pick the right serving path — hosted API for almost everyone, vLLM + DSpark for a real self-hosted run

The one-sentence version​

DeepSeek V4-Flash-Vision-Exp is a 305-billion-parameter sparse-MoE multimodal model — 13B active per token, 1.048M-token context — that adds joint-trained vision to the V4-Flash text backbone, ships open under MIT on Hugging Face on August 31 2026, and delivers agent-tier multimodal performance at roughly one-fortieth of Claude Opus 4.8's per-token price.

Four things about V4-Flash-Vision that aren't in the release notes​

1. Vision was added without breaking the text agents — that's the whole point​

Every open lab that has bolted vision onto a strong text model has paid a tax on text benchmarks. DeepSeek's release table is the flex: on Terminal Bench 2.1 the vision model scores 83.9 (vs 82.7 for the text-only V4-Flash-0731); on DeepSWE it climbs to 59.3 (from 54.4); on Toolathlon-Verified to 75.9 (from 70.3). None of these are vision tasks — they're pure code and tool-use agent evals. The vision variant is better on text-agent work than the text-only checkpoint, not worse.

This matters because it's the "prove-out" the field has been waiting for since Chameleon: you can jointly train a giant MoE for text and vision without the multimodal add causing text regression, as long as the vision head lands late enough and the router gets to keep specializing text experts.

2. The vision encoder has a hard 384-tokens-per-image budget​

This is the single most important thing to learn before writing your first prompt. The vision encoder in V4-Flash-Vision-Exp allocates at most 384 tokens per image — and that ceiling is architectural, not a commercial policy. There is no "high-detail" mode that unlocks more. Practical consequences:

  • Big screenshots lose fine text. A 2560×1440 dashboard, a dense CSV rendered as an image, or a long PDF page will be encoded at whatever resolution 384 tokens of visual features can express — meaning small-font text will not survive round-trip.
  • Crop, don't scale. If you need to read numbers off a chart, send the chart cropped to the region of interest, not the whole page.
  • Multi-image beats one giant image. Split large documents into per-region tiles; each tile gets its own 384-token budget.

This is why Chartography lands at 64.3 (competitive with Claude Opus 4.8's 65.0) but why the model will still miss a footnote in a 12-point font on a full-page screenshot. Every vision-agent workflow you build should assume this ceiling.

3. The MoE routing is unusually sparse — 6 of 256 experts fire per token​

Most published MoE recipes activate 8 experts of 32 or 64. V4-Flash-Vision-Exp uses a much wider expert bank: 256 routed experts + 1 shared expert, with only 6 routed experts firing per token — a 2.3% activation ratio. The model has 43 layers, hidden size 4,096, and the hybrid attention stack combines two mechanisms:

  • Compressed Sparse Attention (CSA) — sparse patterns for long-range reads.
  • Heavily Compressed Attention (HCA) — an aggressively compressed form for the deep-context tail.

Wrapped over that, Manifold-Constrained Hyper-Connections (mHC) replace the standard residual with a constrained-manifold projection, which the paper reports stabilizes signal propagation through the 43-layer stack. Together these are the reason the model can advertise a 1,048,576-token context window without the quadratic attention bill.

4. Pricing is the actual story​

Sticker prices, all per million tokens:

ModelInputOutputCached read
DeepSeek V4-Flash-Vision-Exp (API)$0.22$0.66$0.007
Claude Opus 4.8$15.00$75.00$1.50
Kimi K3 (cache-miss / cache-hit)$3.00 / $0.30$15.00$0.30

Cached-read at $0.007/M is not a typo — that is roughly one two-hundredth of a cent per thousand tokens. For any workload where a repeating system prompt or long context is re-sent (agent loops, RAG over a static corpus, chat with a fixed persona) the effective input price collapses. This is the reason V4-Flash-Vision-Exp is showing up in agent-cost tables at a fraction of what Claude Opus 4.8 costs per task, even when the raw quality is within a few points.

Benchmarks — read the table, not the headline​

BenchmarkV4-Flash-Vision-ExpV4-Flash-0731Claude Opus 4.8
Text agents
Terminal Bench 2.183.982.785.0
NL2Repo57.754.269.7
DeepSWE59.354.458.0
Toolathlon-Verified75.970.376.2
Multimodal agents
ApexBench (Pass@1)36.526.2†39.4
Agents' Last Exam27.325.2†25.7
Chartography64.3—65.0
ZeroBench (Pass@5)35.0—34.0

† Prior text-only checkpoint scored on images it couldn't natively read — a floor, not a fair fight.

Reading the table straight:

  • On text-agent work, V4-Flash-Vision-Exp is within a few points of Claude Opus 4.8 across the board, and actively beats V4-Flash-0731 on every eval. Vision was added without regression.
  • On multimodal agents, it beats Claude Opus 4.8 on Agents' Last Exam and ZeroBench Pass@5, trails narrowly on ApexBench and Chartography. This is the first open-weight release where "matches Opus on multimodal agents" is not a marketing line.
  • The one benchmark that's not close is NL2Repo (57.7 vs 69.7). Whole-repository code translation is still a place where the closed frontier holds a real gap.

The release pattern — API first, weights ten days later​

V4-Flash-Vision-Exp launched on the DeepSeek API on August 21, 2026, and only landed on Hugging Face as open weights on August 31. This ten-day gap is DeepSeek's new default. Compare Z.ai (GLM), which has been shipping API + weights simultaneously; DeepSeek is deliberately sequencing.

The industry read: labs now agree that big open MoEs are shipping. The remaining question is when in the release cycle the weights arrive. For production planning that means:

  • If you need on-prem from day zero, watch for GLM-family releases.
  • If you can start on the API and cut over to self-hosted a week or two later, DeepSeek is the pattern.

Running V4-Flash-Vision-Exp​

There are three realistic paths. Pick by the two constraints that actually matter: whether the weights need to live on your infrastructure, and how much VRAM (or RAM) you can pay for.

Guided walkthrough1 of 3
  1. OpenAI-compatible endpoint at api.deepseek.com. Model ID is deepseek-v4-flash-vision-exp. Pricing is $0.22 / $0.66 / $0.007 per million tokens. This is the right answer unless you have a data-residency or air-gap constraint.

Hosted API — Python (OpenAI-compatible)

import openai

client = openai.OpenAI(
  api_key="YOUR_DEEPSEEK_KEY",
  base_url="https://api.deepseek.com/v1",
)

response = client.chat.completions.create(
  model="deepseek-v4-flash-vision-exp",
  messages=[
      {
          "role": "user",
          "content": [
              {"type": "text", "text": "What does this chart claim, and what's the one thing it hides?"},
              {"type": "image_url", "image_url": {"url": "https://example.com/chart.png"}},
          ],
      }
  ],
)
print(response.choices[0].message.content)

Self-host — vLLM on 4× GB300 with DSpark

docker run --gpus all -p 8000:8000 \
vllm/vllm-openai:deepseekv4-flash-vision \
deepseek-ai/DeepSeek-V4-Flash-Vision-Exp \
--kv-cache-dtype fp8 \
--block-size 256 \
--tensor-parallel-size 4 \
--tool-call-parser deepseek_v4 \
--speculative-config '{"method":"dspark","model":"deepseek-ai/DeepSeek-V4-Flash-Vision-Exp","num_speculative_tokens":3}'

Local — Ollama with the community GGUF

ollama run hf.co/unsloth/DeepSeek-V4-Flash-Vision-Exp-GGUF:UD-Q4_K_XL

# Compact image syntax the model was trained to accept:
# <image>path/to/screenshot.png</image>  followed by your prompt

When to reach for V4-Flash-Vision — and when to stay on Claude​

Reach for V4-Flash-Vision when:

  • You need multimodal agent quality at roughly Opus-4.8 tier and cost is a factor. Cached-read pricing crushes anything on the closed frontier.
  • The workload benefits from a 1M-token context (long-document VQA, whole-video-frame analysis, big-repo review with screenshots).
  • You need open weights for compliance, air-gapped work, or fine-tuning, and you have (or rent) frontier-class hardware.
  • You're building a chart-heavy or dashboard-parsing pipeline where you can control image tiling.

Stay on Claude Opus 4.8 (or Sonnet 5) when:

  • The job is whole-repository code translation or synthesis — NL2Repo is the outlier where the gap is real (57.7 vs 69.7).
  • You need Claude-specific features: Extended thinking under budget, Skills, Sub-agent orchestration in Claude Code, or the Managed Agents session-budget controls.
  • Your prompts contain images with tiny critical text and you cannot tile — you'll hit the 384-token cap and lose accuracy.
  • You want a single vendor / SLA / support contact and are already invested in Anthropic tooling.

For a broader model-choice framework, see How to choose a model. For a side-by-side of the whole open-weight wave, DeepSeek, Qwen & the Open-Weight Wave is the durable overview.

The gotchas that will bite you first​

  1. You will forget the 384-token cap. The first three vision agents you build will overshoot resolution and quietly lose accuracy on small text. Instrument for it: log per-image dimensions and downscale + tile deliberately.
  2. DSpark drops silently if you skip --speculative-config. Without the flag, vLLM will happily serve the model at 1/3 the throughput the release announced. Always pass the speculative config in production.
  3. The GGUF is not a laptop model. UD-Q4_K_XL is 155 GB. Community threads that show it "running" on smaller boxes are running it at fractional tokens/s via SSD swap — technically working, not usable.
  4. API vs weights version parity is not guaranteed. The API model can be silently updated; the Hugging Face weights are pinned to Vision-Exp. If a benchmark shifts a few points month-to-month on the API, that's why.
  5. Exp means experimental. The vision variant is not guaranteed to survive into production V4 releases. Do not build a critical pipeline on the assumption that this specific checkpoint will be around in a year.

Quiz​

Check yourself

0/5
  1. How many parameters activate per token in DeepSeek V4-Flash-Vision-Exp?
  2. What is the hard architectural limit on how many tokens the vision encoder allocates to a single image?
  3. On which benchmarks in the release table does V4-Flash-Vision-Exp actually beat Claude Opus 4.8?
  4. What is the cached-read input price on the DeepSeek API for this model?
  5. Which single-node hardware profile is the vLLM-recommended path for self-hosting V4-Flash-Vision-Exp in production?
Press Enter or Space to flip the card. Use the left and right arrow keys to move between cards.Term shown.
1 / 11

Sources & further reading​

Next​