AI CRITIQUE

DeepSeek logosu, markanın adını ve dinamik, soyut bir ışık etkisi üzerine bindirilmiş stilize bir balina grafiğini içerir.

DeepSeek unveils new vision model with performance close to Claude Opus 4.8

Read the article English (automatic translation)

Summary

DeepSeek has released DeepSeek-V4-Flash-Vision-Exp, an experimental vision-capable variant of its V4 Flash model. The company claims the model performs near Anthropic's Claude Opus 4.8 in visual interpretation. It maintains the low-price policy of its text-only sibling model. Accessible via API, the model supports mixed text and image inputs and is currently free for testing.

Why it matters

Adding multimodal vision to a model that achieves near-frontier reasoning capabilities at a fraction of the cost of Western rivals narrows the cost-performance gap that has been the key differentiator for Claude and GPT-class products. If the benchmark claims are verified, this puts pressure on both Anthropic and OpenAI on both price and capability. It also intensifies the competitive pressure already seen with Kimi K3 and Qwen 3.8. For developers building agent stacks, a free, API-native vision model that integrates with mainstream frameworks lowers the barrier to launching multimodal products.

Highlights

Vision capabilities were added to V4 Flash without additional cost to the text layer; billing limits each image to 384 tokens
Self-reported performance in visual reasoning that 'approaches Claude Opus 4.8'; capable of executing agent-style tasks without continuous oversight
API access only; supports base64, URL, or file upload inputs; allows up to 600 images per request and supports four common image formats (JPEG, PNG, GIF, WebP)

Critical take

The 'approaching Opus 4.8' framing is entirely DeepSeek's own marketing claim. The article cites no independent benchmarks or reproducible evaluations. The model is also explicitly labeled 'experimental,' a tag that typically indicates instability or limited readiness for production environments. The 384-token limit per image and the restriction to only four raster formats (excluding PDFs, video, and structured documents) suggest that the vision stack is still a thin add-on rather than a full multimodal system.

Bu kritik, yerel yapay zeka modeli Qwen3.8-27B tarafından yazılmıştır. Haberi ben seçtim; kritik metnini yayımlanmadan önce tek tek denetlemiyorum.This critique was written by the local AI model Qwen3.8-27B. I chose the story; I do not review each critique before it is published.

Back to the issue