Local Agent Models and Open Agent Tools #99
Today's Letter
- LiquidAI, LFM2.5-2.6B local agent model released
- Microsoft Research, Orchard open agent framework released
- Vercel, browser tools added to eve agents
LiquidAI, LFM2.5-2.6B local agent model released
- LiquidAI released LFM2.5-2.6B for on-device agents with tool calling and multi-step workflow support
- The model is designed for laptops and phones, with inference under 2.5 GB of memory
- Reported decode speed reaches 220 tok/s on Apple M5 Max and 113 tok/s on AMD Ryzen AI Max+ 395
- Training used approximately 34T tokens, extended the context window to 128K, and included agentic reinforcement learning
- LFM2.5-2.6B leads the listed instruction-following benchmarks and most tool-use benchmarks against models up to roughly 4x larger
- It supports llama.cpp, MLX, vLLM, SGLang, ONNX, and Transformers 5.0.0 or newer
- LFM2.5-2.6B and LFM2.5-2.6B-Base are available on Hugging Face, with a WebGPU browser demo
Source: huggingface.co
Microsoft Research, Orchard open agent framework released
- Microsoft Research released Orchard, an open-source framework for scalable and cost-effective agentic AI research
- Orchard Env provides reusable Kubernetes-based environments for training, reinforcement-learning rollouts, and evaluation
- The same infrastructure supports software-engineering, web-navigation, and personal-assistant agents
- Agents can be trained inside deployment harnesses including Codex, OpenClaw, and ZeroClaw through isolated containerized rollouts
- Orchard-SWE reaches 69.7% on SWE-bench Verified, or 73.0% with value-model reranking, using about 3 billion active parameters
- The release includes Orchard-SWE, Orchard-GUI, Orchard-Claw, training data, and evaluation methods
Source: microsoft.com
More: opensourceforu.com
Vercel, browser tools added to eve agents
- eve agents can now navigate the web through the @agent-browser/eve extension
- Available tools include navigation, content reading, clicks, form filling, screenshots, and console and network inspection
- Chromium and agent-browser run inside the eve sandbox, isolated from the application
- allowedDomains restricts navigation and subresources by domain
- Extension options control output size, screenshot handling, proxy use, and session naming
- Cookie, storage, and saved-auth-state commands are not exposed to the model
- Install with pnpm add @agent-browser/eve and mount the extension under agent/extensions/
Source: vercel.com
More: github.com
Jocoletter curates AI, software, and product trends for developers and builders.
#LiquidAI #Microsoft #Vercel