DeepSeek unveils an experimental multimodal version of its V4 Flash model, saying it nears the performance of Anthropic's Opus 4.8 on multimodal agentic tests
Bloomberg
Context & Ripple Effects
DeepSeek's 2026 cadence has been a steady climb up the frontier: the April preview conceded V4 Pro trailed state-of-the-art models by three to six months, then the official V4 Flash API entered public beta in late July with agent-focused claims, and by August 1 V4 Flash hit 50 on the Artificial Analysis Intelligence Index, matching Gemini 3.6 Flash.
Today's move extends that text-model progress into multimodal agentic territory — the workload Anthropic's Opus 4.8 defines — and it lands on infrastructure DeepSeek prepared early: reporting from February said it worked with Huawei and Cambricon to optimize the multimodal V4 for their chips.
First-order effects
Anthropic's Opus 4.8 is now explicitly the yardstick DeepSeek measures against on multimodal agentic tests, though the 'nears' framing is DeepSeek's own claim rather than an independent benchmark result.
Buyers evaluating agents for vision-plus-tool workflows gain a second credible candidate alongside Gemini 3.6 Flash, which V4 Flash already matches on the Artificial Analysis index.
Second-order effects
If the multimodal variant follows V4 Flash's pricing posture into the public-beta API, Anthropic and Google face margin pressure on exactly the agentic workloads they currently price as premium.
The Huawei/Cambricon optimization path means DeepSeek's capability gains are decoupled from export-controlled accelerators, weakening the assumption that chip access gates frontier parity.
Third-order effects
DeepSeek's own April admission that its flagship trailed the frontier by three to six months is becoming the template: frontier gaps measured in quarters rather than years, forcing US labs to compete on distribution and enterprise trust instead of raw capability leads.
Agentic benchmarks are consolidating as the arena where Chinese and American labs contest leadership, shifting the comparison point from chat quality to autonomous task completion.
The trend: Chinese labs are compressing the frontier capability gap to a few months by iterating fast on domestically optimized silicon, turning agentic benchmarks into the new battleground with US leaders like Anthropic.
DeepSeek-V4-Flash-Vision-Exp is now live on the DeepSeek API Platform! 🚀 🔹 This experimental multimodal model matches DeepSeek-V4-Flash on text capabilities—including agents, reasoning, and world knowledge. 🔹 On multimodal agent benchmarks, V4-Flash-Vision-Exp makes a major
Multimodal API support 🔌 🔹 Set model='deepseek-v4-flash-vision- exp' 🔹 Images are tokenized for billing: up to 384 tokens each, at V4-Flash pricing 🔹 Supports Chat Completions, Messages & Responses 🔹 Supports mixed text + image input; images can be provided via base64, external U…
DeepSeek V4 Flash Vision Exp is live on OpenRouter! @deepseek_ai's new model supports Image input at V4 Flash pricing, matches V4 Flash 0731 on text agent benchmarks, and beats Opus-4.8 on Agents' Last Exam and ZeroBench. Use it now: https://openrouter.ai/...
I will move my low cost AI task to this new flash version. In just half a month, it got event better at SWE + added vision support. Opus4.8 level SWE @ DeepSeek flash price. Great stuff.
wtf is happening today: DeepSeek has launched DeepSeek-V4-Flash-Vision-Exp, an experimental multimodal model built for agents that need to see. And its performance on visual-agent benchmarks moves close to or even outperforms Opus 4.8. Again: this is the Flash model, the small …
This is important too Two more notes: - they don't call models “exp” or “preview” as a joke. 3.2-exp was half a generation behind 3.2. Vision-Full will go crazy - nothing said on open weights so far
Files API is now live. 📁 🔹 Free to use 🔹 Upload an image once, then reference it by file_id to save request bandwidth 🔹 Reuse the same image across requests—no need to upload it again Learn more: https://api-docs.deepseek.com/ ... 4/n
Multimodality unlocks more agent use cases. 👀 V4-Flash-Vision-Exp works smoothly across agent frameworks, combining visual understanding with a wide range of tools to unlock more practical workflows. 2/n