Here is your today’s AI Dev Brief from Marktechpost, covering core research, models, infrastructure tools, and applied updates for AI developers and researchers. BTW, we just released our Latest Machine Learning Global Impact Report along with Interactive Graphs and Data: Revealing Geographic Asymmetry Between ML Tool Origins and Research Adoption. Feel free to check out the platform, its completely free and no login required.
OpenAI has Released the ‘circuit-sparsity’: A Set of Open Tools for Connecting Weight Sparse Models and Dense Baselines through Activation Bridges
The release provides a pretrained transformer on Hugging Face under the name openai/circuit-sparsity together with a public GitHub repository openai/circuit_sparsity that contains model checkpoints, Python task definitions and a browser based circuit explorer. The model is trained with enforced weight sparsity during optimization so that each parameter matrix contains only a small fraction of nonzero entries and activations are also encouraged to be sparse, which leads to compact and more isolated computation paths inside the network. On top of this model, the toolkit exposes fully instantiated circuits for concrete Python coding behaviors such as choosing the correct closing quote for strings, deciding between single and double bracket closures based on nesting depth. Read the full insights/article here.
CopilotKit v1.50 Brings AG-UI Agents Directly Into Your App With the New useAgent Hook
Agent frameworks are now good at reasoning and tools, but most teams still write custom code to turn agent graphs into robust user interfaces with shared state, streaming output and interrupts. CopilotKit targets this last mile. It is an open source framework for building AI copilots and in-app agents directly in your app, with real time context and UI control. The release of of CopilotKit’s v1.50 rebuilds the project on the Agent User Interaction Protocol (AG-UI) natively. The key idea is simple; Let AG-UI define all traffic between agents and UIs as a typed event stream to any app through a single hook, useAgent.....
Nanbeige4-3B-Thinking: How a 23T Token Pipeline Pushes 3B Models Past 30B Class Reasoning
Nanbeige LLM Lab at Boss Zhipin released Nanbeige4-3B-Thinking-2511, a 3B SLM pretrained on 23T high quality tokens and post trained with 30M plus instructions, using FG-WSD curriculum scheduling, Dual-Level Preference Distillation, and multi stage GRPO RL, and it posts AIME 2024 avg@8 90.4 and GPQA-Diamond avg@3 82.2, exceeding Qwen3-32B-2504 on AIME 2024 at 81.4 and Qwen3-14B-2504 on GPQA-Diamond at 64.0, while still trailing larger models on some coding heavy benchmarks like Fullstack-Bench. Read the full insights/article here.
Thinking Machines Lab Makes Tinker Generally Available: Adds Kimi K2 Thinking And Qwen3-VL Vision Input
Tinker from Thinking Machines Lab is now generally available, with support for fine tuning Moonshot AI’s 1 trillion parameter Kimi K2 Thinking reasoning model, OpenAI API compatible inference on in training checkpoints, and vision input through Qwen3-VL models, so AI engineers can run research grade customization of large reasoning and multimodal models through a simple Python training loop without building or maintaining their own distributed training stack. Read the full insights/article here.
Project Notebooks/Tutorials
▶ [Open Source] Rogue: An Open-Source AI Agent Evaluator worth trying Codes & Examples
▶ How to Design a Fully Local Agentic Storytelling Pipeline Using Griptape Workflows, Hugging Face Models, and Modular Creative Task Orchestration Codes Tutorial
▶ How to Design a Gemini-Powered Self-Correcting Multi-Agent AI System with Semantic Routing, Symbolic Guardrails, and Reflexive Orchestration Codes Tutorial
▶ A Coding Implementation of Secure AI Agent with Self-Auditing Guardrails, PII Redaction, and Safe Tool Access in Python Codes Tutorial
▶ Meet LangChain’s DeepAgents Library and a Practical Example to See How DeepAgents Actually Work in Action Codes Tutorial