We read a lot of papers. Most are incremental. A few change how we think about what is possible. Here are three from the second half of 2025 that have directly influenced how we build at [r]think and deetech.

We have included local copies of each paper so you can read them offline.

1. DINOv3: A Foundation Model for Unifying Vision

Meta FAIR, August 2025 · Read PDF · arXiv 2508.10104

DINOv3 is Meta's 7B parameter self-supervised vision foundation model, trained on 1.7 billion images. The headline result: for the first time, a purely self-supervised model beats weakly supervised approaches across the board on vision tasks -- image classification, object detection, segmentation, depth estimation, and more.

The architecture builds on the DINOv2 line but scales aggressively, combining a ViT-giant backbone with a distillation-based training recipe that extracts robust visual features without any labels. The features transfer directly to downstream tasks with minimal fine-tuning.

Why it matters to us: Vision foundation models are the backbone of any image analysis pipeline, including deepfake detection. DINOv3 proves that self-supervised pre-training at scale produces features that generalise better than supervised approaches. For deetech, this means richer visual representations to build detection on top of -- and the self-supervised nature means we are not limited by labelled deepfake datasets, which are always lagging behind the latest generators.

2. DeepSeek-V3.2

DeepSeek AI, December 2025 · Read PDF · arXiv 2512.02556

DeepSeek-V3.2 is the open-weights model that put frontier performance within reach of everyone. Using sparse attention and scalable reinforcement learning, it achieves performance comparable to GPT-5 class models while remaining open and runnable on reasonable hardware.

The architecture uses a mixture-of-experts design with efficient routing, combined with the reasoning capabilities pioneered in DeepSeek-R1. The training recipe combines supervised fine-tuning with reinforcement learning from verifiable rewards, building on the insight that RL can unlock reasoning capabilities that supervised training alone cannot.

Why it matters to us: Open frontier models change the economics of building AI products. When the base model is free and capable, the competitive advantage shifts entirely to domain expertise, data curation, and system design -- exactly where a focused venture studio has an edge over horizontal players. DeepSeek-V3.2 makes it even clearer that the moat is not the model. The moat is everything around it.

3. Small Language Models are the Future of Agentic AI

NVIDIA Research, September 2025 (revised) · Read PDF · arXiv 2506.02153

This NVIDIA paper makes a sharp argument: large language models are overkill for most agentic tasks. In agentic systems, models perform a small number of specialised tasks repetitively -- routing, tool selection, summarisation, extraction. These tasks do not require general-purpose conversational ability. Models under 10B parameters are not just sufficient, they are inherently more suitable and necessarily more economical.

The authors outline a general LLM-to-SLM conversion algorithm for agent workflows and argue that heterogeneous systems -- agents calling different sized models for different subtasks -- are the natural architecture. The paper has already accumulated over 100 citations.

Why it matters to us: At deetech, our detection pipeline is an agentic system. Images come in, get routed through preprocessing, analysed by detection models, results get formatted for the claims system. Most of these steps do not need a 70B model. They need a fast, reliable, purpose-built model that does one thing well. This paper validates the architecture we are already building: small, specialised models composed into a system, not one monolithic model trying to do everything.

The common thread

All three papers point in the same direction: capability is decoupling from scale. Self-supervised vision beats supervised. Open models match closed frontier. Small models outperform large ones at specialised tasks. You do not need the biggest model or the most compute to build products that work.

For a venture studio building AI products, this is the most important structural shift in the landscape. It means small teams with domain expertise and engineering discipline can build products that compete with (or outperform) what well-funded horizontal players build with brute force. That is the thesis we are building on.