Here is your today’s AI Dev Brief from Marktechpost, covering core research, models, infrastructure tools, and applied updates for AI developers and researchers.
From Gemma 3 270M to FunctionGemma, How Google AI Built a Compact Function Calling Specialist for Edge Workloads
FunctionGemma is Google’s 270M parameter variant of Gemma 3 that is trained specifically for function calling, so it turns natural language instructions into structured API or tool calls on device rather than acting as a general chat model. It keeps the Gemma 3 architecture and 256K token vocabulary, supports 32K tokens per request, and is trained on 6T tokens with an August 2024 knowledge cutoff, which makes it compact enough for phones, laptops and Jetson class hardware. The model uses a strict chat template with dedicated control tokens for function declarations, function calls and function responses, and benchmarks like Mobile Actions show accuracy improving from 58 percent to 85 percent after task specific fine tuning, which highlights how small edge agents become reliable when you specialize them on your own tools and workflows. Read the full insights/article here.
Liquid AI’s LFM2-2.6B-Exp Uses Pure Reinforcement Learning RL And Dynamic Hybrid Reasoning To Tighten Small Model Behavior
Liquid AI has released LFM2 2.6B Exp, an experimental 3B class language model that adds a pure reinforcement learning stage with verifiable rewards on top of the pretrained and preference aligned LFM2 2.6B backbone, targeting instruction following, knowledge tasks, and math while keeping an edge friendly footprint. The model uses a hybrid architecture with short range LIV convolutions and grouped query attention, a 32,768 token context and a 10 trillion token training budget, and it reports IFBench scores that exceed those of much larger models like DeepSeek R1 0528, making it a strong candidate for agentic systems, structured extraction, and on device assistants. Read the full analysis….
NVIDIA AI Researchers Release NitroGen: An Open Vision Action Foundation Model For Generalist Gaming Agents
NVIDIA AI Researchers Release NitroGen: An Open Vision Action Foundation Model For Generalist Gaming AgentsNitroGen is an open vision action foundation model from NVIDIA and collaborators that learns to play commercial video games directly from pixels and standardized gamepad actions, trained purely with large scale behavior cloning on 40,000 hours of gameplay videos from more than 1,000 titles where player inputs are recovered from on screen controller overlays, and shipped with an internet scale action labeled dataset, a Gymnasium compatible universal simulator for Windows games, and a 4.93 × 10^8 parameter SigLIP 2 plus diffusion transformer policy that achieves strong multi game performance and up to 52 percent relative improvement in task success when fine tuned on unseen games compared to training from scratch. Read the full analysis….
Project Notebooks/Tutorials
▶ [Open Source] Rogue: An Open-Source AI Agent Evaluator worth trying Codes & Examples
▶ How to Build Production-Grade Agentic Workflows with GraphBit Using Deterministic Tools, Validated Execution Graphs, and Optional LLM Orchestration Codes Tutorial
▶ A Coding Guide to Build an Autonomous Multi-Agent Logistics System with Route Planning, Dynamic Auctions, and Real-Time Visualization Using Graph-Based Simulation Codes Tutorial