Liquid AI Launches LocalCowork: Privacy-First Agent Workflows with LFM2-24B-A2B

Liquid AI Releases LocalCowork Powered By LFM2-24B-A2B to Execute Privacy-First Agent Workflows Locally Via Model Context Protocol (MCP)

Liquid AI has released LFM2-24B-A2B and the LocalCowork desktop agent to enable privacy-first workflows on local hardware. The system achieves sub-second tool dispatch, averaging approximately 385 ms per response on an Apple M4 Max.

Why This Matters

While cloud-based LLMs offer high reasoning capabilities, they introduce data egress risks and latency overhead unsuitable for privacy-sensitive enterprise environments. LocalCowork addresses this by running entirely on-device, though technical benchmarks reveal a performance gap in complex tasks, with multi-step chain completion dropping to 26% compared to 80% for single-step actions. This underscores the current technical reality where local models are highly efficient for discrete tasks but require human-in-the-loop oversight for complex autonomous operations.

Key Insights

• Sparse Mixture-of-Experts (MoE) design in LFM2-24B-A2B activates only 2 billion of its 24 billion parameters per token, reducing computational overhead for local inference (Liquid AI, 2026).
• Quantization in Q4_K_M GGUF format allows the 24B model to fit into a ~14.5 GB RAM footprint, enabling execution on consumer-grade Apple M4 Max hardware.
• The Model Context Protocol (MCP) facilitates integration with 75 tools across 14 servers, handling tasks like filesystem operations and OCR without cloud API dependencies.
• Performance testing indicates a 26% success rate on multi-step chains due to ‘sibling confusion,’ where the model selects similar but incorrect tools during complex logic.
• LocalCowork maintains a local audit trail for compliance, recording every tool call locally to meet strict data privacy and regulatory requirements.

Practical Applications

• Security Scanning: Identifying leaked API keys and PII within local directories while maintaining a zero-data-egress posture to avoid exposing sensitive credentials to third-party APIs.
• Document Processing: Executing OCR and parsing contracts locally to generate PDFs; however, users must watch for ‘sibling confusion’ where the model might pick an incorrect parsing tool during multi-step analysis.

References:

• https://www.marktechpost.com/2026/03/05/liquid-ai-releases-localcowork-powered-by-lfm2-24b-a2b-to-execute-privacy-first-agent-workflows-locally-via-model-context-protocol-mcp/