DeepSeek Releases New Experimental Model, Claims Near-Parity With Anthropic’s Opus 4.8
The model accepts image and text inputs and is designed to support visual understanding within agentic workflows.
Chinese AI company DeepSeek has launched an experimental multimodal model designed to analyse images, screenshots and visual prompts while retaining the text reasoning and agentic capabilities of its V4-Flash model, Bloomberg reported.
The DeepSeek-V4-Flash-Vision-Exp is now available through DeepSeek’s API. The company claims the model delivers multimodal agentic performance close to Anthropic’s Opus 4.8, putting the experimental release directly into competition with one of the leading models for complex AI tasks.
The model accepts image and text inputs and is designed to support visual understanding within agentic workflows. DeepSeek says images can be submitted through base64 encoding, external URLs or its Files API, with images converted into tokens for processing and billing.
The release comes as DeepSeek reportedly prepares for a potential public listing. Bloomberg has reported that the Hangzhou-based company is preparing for a mainland China IPO, potentially filing in 2026 ahead of a possible 2027 debut.
DeepSeek is also reportedly seeking additional private funding at a valuation of at least 480 billion yuan ($71 billion). The potential round would follow a reported $7 billion financing and value the company significantly above the roughly $50 billion valuation associated with its first external funding round, which included Tencent and Contemporary Amperex Technology.
DeepSeek’s rapid model evolution
DeepSeek's rapid rise has been driven by a succession of increasingly capable models. The company began releasing its DeepSeek-LLM and DeepSeek Coder models in late 2023, followed by DeepSeek-MoE and DeepSeek-Math in 2024. DeepSeek-V2 arrived in May 2024, followed by Coder-V2 and V2.5.
Its biggest breakthrough came with DeepSeek-V3, released in December 2024. The model used a 671-billion-parameter mixture-of-experts architecture, with 37 billion parameters activated for each token.
DeepSeek then released R1 in January 2025, a reasoning-focused model that rapidly gained global attention and challenged assumptions about the computing resources required to build competitive AI systems.
The company subsequently released updated reasoning and general-purpose models, including R1-0528, V3.1 and V3.2, before introducing the V4 family in 2026.
The new Vision-Exp model extends that progression into multimodal AI, signalling DeepSeek’s push beyond text-based reasoning toward agents capable of interpreting and acting on visual information.
