llms.blog
AI

DeepSeek Unveils Experimental Vision Model Challenging Anthropic's Opus 4.8

DeepSeek Unveils Experimental Vision Model Challenging Anthropic's Opus 4.8 DeepSeek announced an experimental multimodal version of its V4 Flash model that can analyze visual prompts, claiming near-parity with Anthropic's Opus 4.8 on multimodal agentic benchmarks. The new release, deepseek-v4-flash-vision-exp, extends DeepSeek's flagship text-only V4 Flash model with vision capabilities. The experimental model processes images alongside text, enabling use cases like describing pictures, rea

axc
axc
2 min
DeepSeek Unveils Experimental Vision Model Challenging Anthropic's Opus 4.8

DeepSeek Unveils Experimental Vision Model Challenging Anthropic's Opus 4.8

Illustration: Neural network forming vision patterns

DeepSeek announced an experimental multimodal version of its V4 Flash model that can analyze visual prompts, claiming near-parity with Anthropic's Opus 4.8 on multimodal agentic benchmarks.

The new release, deepseek-v4-flash-vision-exp, extends DeepSeek's flagship text-only V4 Flash model with vision capabilities. The experimental model processes images alongside text, enabling use cases like describing pictures, reading text from screenshots, and analyzing charts.

How the Model Works

DeepSeek treats images as tokens for billing purposes. Each image is converted into up to 384 tokens, billed at V4-Flash pricing. The model accepts JPEG, PNG, GIF, and WebP formats, with images resized during processing to maintain consistent token budgets.

Developers can provide images via:

  • Base64-encoded inline data
  • External HTTP URLs
  • Files API uploads (allowing images up to 64 MiB)

The API supports detail levels (low, high, original) and integrates with OpenAI-compatible endpoints, the Anthropic-compatible /messages API, and the Responses API for agent workflows.

Benchmark Claims

According to DeepSeek's X announcement, V4-Flash-Vision-Exp "moves close to or even outperforms Opus 4.8" on multimodal agent benchmarks including Agents' Last Exam and ZeroBench. The company positions the model as a lower-cost alternative for vision-enabled agent tasks.

Availability

The model launched on DeepSeek's API platform with vision support enabled at V4 Flash's pricing tier ($0.002 per 1K tokens input, $0.006 per 1K tokens output). It also launched on OpenRouter with the same pricing.

DeepSeek describes the release as experimental, indicating further tuning and scaling work ahead.

Strategic Context

The release comes amid DeepSeek's aggressive push in the AI model market. In July 2026, the company reported its Claude Code competition efforts, and in August it announced plans for an IPO. The vision model extends DeepSeek's multimodal capabilities, which previously included V1 and V2 versions in 2023-2024.

Sources

AIDeepSeekMultimodalVisionAnthropic
ShareX (Twitter)LinkedIn

Written by

axc

More to read

  • Sharpness-Aware Minimization in Large Language Models: How Adversarial Weight Perturbations and Flat Minima Boost Generalization

    In overparameterized deep neural networks, minimizing empirical training loss is insufficient to guarantee optimal generalization on unseen distributions. Modern deep architectures, including vision models and autoregressive Large Language Models (LLMs), operate in regimes where parameter counts far exceed training token counts, producing highly non-convex loss surfaces populated by infinite global minima. Standard optimization via Stochastic Gradient Descent (SGD) or AdamW often converges to sh

    1 min

  • Prefix-Tree KV Cache Management in Production: RadixAttention, Tree-Structured LRU Eviction, and Token-Level Sharing in SGLang and vLLM

    Prefix-Tree KV Cache Management in Production: RadixAttention, Tree-Structured LRU Eviction, and Token-Level Sharing in SGLang and vLLM Autoregressive large language model inference is heavily constrained by memory bandwidth and the computational overhead of the prefill phase. For workloads such as multi-turn conversations, autonomous agent tool loops, few-shot prompt evaluations, and tree-search decoding, consecutive requests often share substantial token prefixes. In a standard multi-turn ses

    1 min

  • Neural Collapse: How Simplex Equiangular Tight Frames Emerge at the Terminal Phase of Training

    In classification tasks, deep neural networks exhibit an unexpected geometric simplicity during late-stage optimization. While the internal activations of early training appear high-dimensional and complex, the penultimate layer representations and linear classifiers converge toward an exact, symmetrical geometric structure known as Neural Collapse (NC). First identified empirically by Papyan, Han, and Donoho (2020), Neural Collapse emerges during the Terminal Phase of Training (TPT). This regi

    1 min