The landscape of AI agents has been dominated by large language models (LLMs) like GPT-4 and Claude, but a new frontier is opening up: lightweight, open-source, locally-deployable agents that can run on consumer hardware. This post shares internal notes and discoveries from my journey building agents for small language models (SLMs) – models ranging from 270M to 32B parameters that run efficiently on CPUs or modest GPUs. These are lessons learned from hands-on experimentation, debugging, and optimizing inference pipelines.
SLMs offer immense potential: privacy through local deployment, predictable costs, and full control thanks to open weights. However, they also present unique challenges that demand a shift in how we design agent architectures. — Read More
Daily Archives: August 28, 2025
InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency
We introduce InternVL 3.5, a new family of open-source multimodal models that significantly advances versatility, reasoning capability, and inference efficiency along the InternVL series. A key innovation is the Cascade Reinforcement Learning (Cascade RL) framework, which enhances reasoning through a two-stage process: offline RL for stable convergence and online RL for refined alignment. This coarse-to-fine training strategy leads to substantial improvements on downstream reasoning tasks, e.g., MMMU and MathVista. To optimize efficiency, we propose a Visual Resolution Router (ViR) that dynamically adjusts the resolution of visual tokens without compromising performance. Coupled with ViR, our Decoupled Vision-Language Deployment (DvD) strategy separates the vision encoder and language model across different GPUs, effectively balancing computational load. These contributions collectively enable InternVL3.5 to achieve up to a +16.0\% gain in overall reasoning performance and a 4.05x inference speedup compared to its predecessor, i.e., InternVL3. In addition, InternVL3.5 supports novel capabilities such as GUI interaction and embodied agency. Notably, our largest model, i.e., InternVL3.5-241B-A28B, attains state-of-the-art results among open-source MLLMs across general multimodal, reasoning, text, and agentic tasks — narrowing the performance gap with leading commercial models like GPT-5. All models and code are publicly released. — Read More
DINOv3: Self-supervised learning for vision at unprecedented scale
Self-supervised learning (SSL) —the concept that AI models can learn independently without human supervision—has emerged as the dominant paradigm in modern machine learning. It has driven the rise of large language models that acquire universal representations by pre-training on massive text corpora. However, progress in computer vision has lagged behind, as the most powerful image encoding models still rely heavily on human-generated metadata, such as web captions, for training.
Today, we’re releasing DINOv3, a generalist, state-of-the-art computer vision model trained with SSL that produces superior high-resolution visual features. For the first time, a single frozen vision backbone outperforms specialized solutions on multiple long-standing dense prediction tasks including object detection and semantic segmentation. — Read More
China unveils bionic antelope robot to observe endangered Tibetan species
A lifelike robotic Tibetan antelope is now roaming the high-altitude wilderness of Hoh Xil National Nature Reserve in Northwest China’s Qinghai Province.
Equipped with 5G ultra-low latency networks and advanced artificial intelligence (AI) algorithms, the bionic robot is being used to collect real-time data on Tibetan antelope populations without disturbing them.
This is the first time such a robotic antelope has been deployed in the heart of Hoh Xil, which sits more than 15,092 feet (4,600 meters) above sea level. — Read More
CodeMonkeys: Scaling Test-Time Compute for Software Engineering
Scaling test-time compute is a promising axis for improving LLM capabilities. However, test-time compute can be scaled in a variety of ways, and effectively combining different approaches remains an active area of research. Here, we explore this problem in the context of solving real-world GitHub issues from the SWE-bench dataset. Our system, named CodeMonkeys, allows models to iteratively edit a codebase by jointly generating and running a testing script alongside their draft edit. We sample many of these multi-turn trajectories for every issue to generate a collection of candidate edits. This approach lets us scale “serial” test-time compute by increasing the number of iterations per trajectory and “parallel” test-time compute by increasing the number of trajectories per problem. With parallel scaling, we can amortize up-front costs across multiple downstream samples, allowing us to identify relevant codebase context using the simple method of letting an LLM read every file. In order to select between candidate edits, we combine voting using model-generated tests with a final multi-turn trajectory dedicated to selection. Overall, CodeMonkeys resolves 57.4% of issues from SWE-bench Verified using a budget of approximately 2300 USD. Our selection method can also be used to combine candidates from different sources. Selecting over an ensemble of edits from existing top SWE-bench Verified submissions obtains a score of 66.2% and outperforms the best member of the ensemble on its own. We fully release our code and data at https://scalingintelligence.stanford.edu/pubs/codemonkeys/. — Read More
Rick's Cafe AI 