88Z AI, 3 May — DeepSeek has released a multimodal large model and its technical report on GitHub, introducing a thinking framework based on visual primitives. By converting spatial markers into basic reasoning units, the model significantly enhances spatial reference and deduction capabilities. With a compact and efficient architecture, it performs excellently in counting and spatial reasoning tests, rivaling advanced models like GPT-5.4 and Claude-Sonnet-4.6.
At a glance
- Introduces a visual primitive-based thinking framework to enhance spatial reasoning
- Features a compact, efficient architecture with performance rivaling leading models like GPT-5.4
The update
DeepSeek recently officially released its latest multimodal large model on GitHub, accompanied by a detailed technical report. The core of the report introduces an innovative framework called "thinking based on visual primitives." In this framework, spatial markers such as points and bounding boxes are elevated to "basic thinking units" for model reasoning. This design aims to break through the bottlenecks of traditional chain-of-thought approaches in handling complex spatial reference tasks, endowing the model with more precise spatial reference and logical deduction capabilities. In terms of technical implementation, the model adopts a compact architectural design and achieves efficient processing of visual tokens. Benchmark results show that it performs outstandingly in counting accuracy and spatial reasoning capabilities. Its overall performance is comparable to current industry-leading multimodal models such as GPT-5.4 and Claude-Sonnet-4.6, demonstrating strong competitiveness.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- IT之家Source published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.