atomic[.]chat (a desktop app that runs LLMs locally) ran a very revealing comparison for local AI agents, on a MacBook P…

Rohan Paul Twitter · Rohan Paul (@rohanpaul_ai) · 2026-05-30

A benchmark run by atomic.chat on a MacBook Pro M5 Max shows Liquid's LFM2.5-8B-A1B model outperforming OpenAI's gpt-oss-20b on local AI agent tasks by completing all tool calls, halving runtime, and using only 4.8GB of memory.

Open original ↗

Appears in

Ultra-Low Latency LLM Inference: Benchmarks and Emerging Enterprise Pricing Tier

Extraction

Topics: local-llmai-agentsmodel-benchmarksedge-inference

Claims

Liquid's LFM2.5-8B-A1B completed every required tool call in the benchmark while gpt-oss-20b did not.
LFM2.5-8B-A1B cut runtime by more than half compared to gpt-oss-20b on the same hardware.
LFM2.5-8B-A1B used only 4.8GB of memory, making it viable on consumer hardware.
Smaller, purpose-optimized models can outperform larger general-purpose models on agentic tasks locally.

Key quotes

Liquid's much smaller LFM2.5-8B-A1B beat gpt-oss-20b by finishing every required tool call, cutting runtime by more than half, and using 4.8GB.