DeepSeek Releases Experimental V4-Flash-Vision-Exp, Claims Near-Parity With Opus 4.8 on Agent Benchmarks
DeepSeek has released V4-Flash-Vision-Exp, an experimental multimodal extension of V4-Flash that adds image understanding while preserving text reasoning capabilities. The company claims the model approaches or beats Anthropic's Opus 4.8 on its internal multimodal agent benchmarks.
DeepSeek has released Deepseek-V4-Flash-Vision-Exp, an experimental multimodal model that extends its V4-Flash text model with image understanding. According to DeepSeek, the model nearly matches Anthropic's Opus 4.8 on the company's internal multimodal agent benchmarks — a claim that has not been independently verified.
What's new
V4-Flash-Vision-Exp adds image processing to the existing V4-Flash architecture while, DeepSeek says, preserving the base model's text reasoning and world-knowledge performance. The model is built for visual agent workflows: it can describe images, extract text from screenshots, and analyze diagrams, and DeepSeek says it's designed to work with tool-use across different agent frameworks.
The model supports JPEG, PNG, GIF, and WebP formats. Notably, it determines image format by reading actual file content rather than relying on filenames or declared MIME types, according to DeepSeek's API documentation.
DeepSeek also shipped version 0.1.1 of its Harness framework alongside the model, with built-in support for the new vision capabilities. The model is accessible through OpenAI's Chat Completions and Responses APIs as well as Anthropic's Messages endpoint, easing integration for developers already building on those ecosystems.
Image handling and limits
Developers have three ways to send images: direct Base64 encoding, publicly accessible URLs (up to 32 MiB), or DeepSeek's new Files API, which is free to use. The Files API allows a single upload to be referenced by ID across multiple requests, with a 64 MiB size cap per file.
An optional "detail" field can downscale images to 512 x 512 pixels to save tokens when fine visual detail isn't required. By default, the model normalizes images to roughly 800 x 800 pixels based on aspect ratio before processing. Regardless of original resolution, DeepSeek says each image costs no more than 384 tokens — a fixed, predictable cost structure for image-heavy workloads.
A single request can include up to 600 images. Maximum edge length is 8,192 pixels per side, though that drops to 4,096 pixels once a request includes 15 or more images. Images are restricted to user messages only.
Pricing
DeepSeek says pricing for V4-Flash-Vision-Exp follows existing V4-Flash rates, meaning no separate surcharge for the vision capability itself. Exact per-token input and output pricing for V4-Flash was not specified in DeepSeek's release materials reviewed here; pricing not yet disclosed in this report.
What this means
DeepSeek's move to fold vision into its Flash line — rather than shipping vision as a separate, premium model — signals a push toward multimodal agents as a default feature rather than an add-on. The fixed 384-token-per-image cost and free Files API suggest DeepSeek is optimizing for high-volume, image-heavy agent workloads like screenshot parsing and UI automation, where cost predictability matters more than peak resolution.
The comparison to Opus 4.8 comes from DeepSeek's own internal benchmarks, not third-party evaluation, so the "rivals Opus 4.8" claim should be treated as a company assertion until independent testing confirms it. Given the "Exp" label, this is explicitly a preview rather than a production-ready release, and DeepSeek may adjust capabilities, limits, or pricing before a stable version ships.
Related Articles
DeepSeek Releases V4 Flash Vision Exp, an Experimental Multimodal MoE Model with 1M Context
DeepSeek has released V4 Flash Vision Exp, an experimental vision-enabled variant of DeepSeek V4 Flash 0731 that adds image understanding while matching the base model's text performance. The sparse mixture-of-experts model uses 13B active parameters out of 284B total and supports a 1M token context window.
Qwen Launches Qwen3.8 27B, an Open-Weight Vision-Language Model with 262K Context
Qwen has released Qwen3.8 27B, a 27-billion-parameter dense vision-language model with a 262K token context window, available now via OpenRouter at $0.45 per million input tokens and $3.20 per million output tokens.
DeepSeek to Quadruple API Prices for V4 Pro and V4 Flash Starting August 16
DeepSeek will raise API output token pricing roughly fourfold starting August 16, introducing peak and off-peak rates for its V4 Pro and V4 Flash models. Despite the increase, DeepSeek remains cheaper than competitors like OpenAI's GPT-5.6 Sol and Moonshot's Kimi K3.
Liquid AI Releases LFM2.5-VL-3B, a 3B-Class Vision-Language Model Built for On-Device Deployment
Liquid AI has released LFM2.5-VL-3B, a multimodal upgrade to its LFM2-VL-3B model built for on-device grounding, object detection, and document OCR. The model runs at 228 tokens/sec on an Apple M5 Max and 116 tokens/sec on an AMD Ryzen AI Max+ 395, using under 3.3 GB of memory.
Comments
Loading...