Poolside Launches Free Laguna XS.2 as US Open Source AI Gains Ground

Poolside, a San Francisco-based AI startup founded in 2023, on Tuesday

designed for agentic coding workflows. The release marks a notable entry from a US company into the open source AI space, which has been dominated by Chinese firms like DeepSeek and Xiaomi offering frontier-level performance at significantly lower costs.

The Laguna XS.2 models are optimized for autonomous coding tasks that go beyond simple text generation, including writing code, using third-party tools, and taking actions independently. According to VentureBeat, Poolside also introduced a coding agent harness called “pool” and a web-based development environment named “shimmer” alongside the model release.

Xiaomi Continues Open Source Push with MiMo-V2.5 Models

Xiaomi reinforced its position in open source AI with the

under the MIT License. The models are available on Hugging Face for commercial use without restrictions.

According to Xiaomi’s ClawEval benchmarks, the Pro model leads the open source field with a 63.8% performance score on agentic “claw” tasks. These tasks involve AI systems that can complete work autonomously through third-party messaging apps, including content creation, account management, and scheduling. The models excel in token efficiency, positioning them favorably as services like GitHub Copilot move to usage-based billing models.

Both MiMo versions appear in the top-left quadrant of Xiaomi’s efficiency charts, indicating high performance while using fewer tokens than competing models. This efficiency translates to lower operational costs for enterprises deploying the models in production environments.

Security Concerns Emerge in Open Source AI Ecosystem

Security researchers at Acronis identified malicious activity targeting AI distribution platforms including Hugging Face and ClawHub. The company discovered nearly 600 malicious skills across 13 developer accounts on ClawHub, designed to distribute trojans, cryptominers, and information stealers for Windows and macOS systems.

Two accounts contained the majority of malicious content: hightower6eu with 334 malicious skills and sakaen736jih with 199. The attacks exploit the modular architecture of AI platforms, where community-built extensions can execute external code with elevated privileges.

One identified payload targeting macOS users is the Atomic macOS Stealer (AMOS), distributed through indirect prompt injection techniques. According to SecurityWeek, threat actors embed hidden instructions in resources that AI systems read, causing agents to download and execute malicious code without user awareness.

Cisco Addresses Model Provenance Challenges

Cisco released an open source Model Provenance Kit to help organizations track the lineage and security of third-party AI models. The tool addresses growing concerns about model repositories where millions of models are available but tracking of modifications and verification of developer claims remains inconsistent.

According to Cisco’s announcement, organizations often cannot verify claims about model sources, vulnerabilities, or training biases. This lack of visibility creates security, compliance, and liability risks, particularly when deploying models in customer-facing applications.

The kit helps enterprises trace incidents back to root causes and identify other affected models in their infrastructure. It also addresses licensing and regulatory compliance requirements as governments implement documentation mandates for AI system usage.

RunPod Flash Eliminates Container Dependencies

RunPod launched Flash, an open source Python tool under the MIT license designed to accelerate AI development by eliminating Docker containerization requirements for serverless GPU infrastructure. The platform targets AI agents and coding assistants like Claude Code, Cursor, and Cline, enabling autonomous hardware orchestration with reduced friction.

Flash supports “polyglot” pipelines that route data preprocessing to CPU workers before transferring workloads to high-end GPUs for inference. According to RunPod CTO Brennen Smith in an interview with VentureBeat, the tool aims to “make it as easy as possible to bring together the cosmos of different AI tooling in a function call.”

The platform includes production-grade features such as low-latency load-balanced HTTP APIs, queue-based batch processing, and persistent multi-datacenter storage. RunPod positions Flash as eliminating the “packaging tax” of AI development by removing containerization overhead.

What This Means

The open source AI landscape is experiencing significant momentum, with US companies like Poolside joining Chinese firms in offering competitive alternatives to proprietary models from OpenAI and Anthropic. This trend reflects growing enterprise demand for models that can be modified, deployed locally, and used without usage restrictions.

However, the security incidents at Hugging Face and ClawHub highlight the challenges of maintaining trust in decentralized AI ecosystems. As adoption accelerates, organizations need robust provenance tracking and security validation processes to prevent supply chain attacks through compromised models.

The efficiency gains demonstrated by Xiaomi’s MiMo models and RunPod’s containerless approach suggest that open source solutions are not just matching proprietary performance but potentially offering superior operational characteristics for specific use cases, particularly in cost-sensitive enterprise deployments.

FAQ

What makes Poolside’s Laguna models different from other open source AI models?
Laguna models are specifically optimized for agentic coding workflows, meaning they can autonomously write code, use tools, and take actions beyond simple text generation. They represent one of the first high-performance open source coding models from a US startup.

How do the security risks in open source AI platforms compare to proprietary models?
Open source platforms face unique risks from malicious community contributions, as demonstrated by the 600 malicious skills found on ClawHub. Proprietary platforms have centralized control but may lack transparency about training data and model modifications.

Why are enterprises choosing open source AI models over proprietary alternatives?
Open source models offer licensing flexibility, local deployment options, and often lower operational costs. Models like Xiaomi’s MiMo-V2.5 provide near-frontier performance while allowing modification and commercial use without restrictions.