DeepSeek Multi-Modal Large Model Released: Visual Primitives Framework Solves Spatial Reasoning Challenges

DeepSeek released a multi-modal large model and supporting technical report on GitHub, proposing an innovative reasoning framework based on "visual primitives" to break through the core bottleneck of multi-modal large language models (MLLMs) in spatial reference tasks.

The current mainstream Chain-of-Thought (CoT) reasoning paradigm is limited to the linguistic field. Existing studies mostly focus on improving visual detail recognition to bridge the "perception gap" but ignore the more fundamental "reference gap"—the ambiguity of natural language leads to imprecise guidance for complex spatial layouts, easily breaking the reasoning chain.

The framework transforms spatial markers like points and bounding boxes into "basic thinking units" for reasoning, embedding them into the thinking chain to give the model "referential" ability, which can anchor abstract cognition to specific image coordinates and achieve precise spatial deduction.

Despite its compact size and low image token budget, the model can still match cutting-edge models like GPT-5.4, Claude-Sonnet-4.6, and Gemini-3-Flash in counting and spatial reasoning benchmark tests, pointing the way for efficient System-2 multi-modal intelligence. Previously, DeepSeek had launched an "image recognition mode" with multi-modal recognition capabilities.