Atomic has launched Atomic Agent, an open-source, local-first AI operator that outperformed Hermes in a recent GAIA Level 1 benchmark, scoring 69.8% compared to Hermes’ 58.5%. Both agents utilized the same local Qwen3.6-35B-A3B model, but Atomic Agent displayed superior efficiency, averaging 217 seconds per task versus 351 seconds for Hermes. This performance is part of Atomic’s commitment to transparency, as they published the full test setup and raw artifacts. Atomic Agent is designed for versatility, running multiple models including Qwen, Gemma, and Llama, and is accessible through a simple installation process on macOS, Windows, and Linux.

Qwen: Qwen refers to a family of large language models developed for high-performance reasoning and generation tasks. It was selected as the shared local model in the Atomic Agent versus Hermes benchmark comparison. The choice allowed consistent evaluation of agent capabilities independent of underlying model differences.
Gemma: Gemma is a series of open large language models optimized for efficient local deployment and versatile task handling. Atomic Agent supports running Gemma models via llama.cpp as part of its multi-model compatibility. This integration expands the agent’s flexibility for users preferring different open-source backends.
Llama: Llama encompasses Meta’s series of open large language models widely used for local inference and agentic applications. Atomic Agent incorporates Llama support through llama.cpp to enable broad model choice. The compatibility helps users leverage familiar and performant models within a fully local agent environment.
Atomic: Atomic develops and maintains open-source tools focused on local AI agents that operate with full user control and on-device processing. It recently released Atomic Agent as its flagship operator designed for privacy-focused automation. The launch positions Atomic as a key player in local-first AI tooling by enabling seamless integration with various models and external services.
Hermes: Hermes is an AI agent designed for task execution and benchmarking in agentic workflows. It was directly compared to Atomic Agent in a standardized GAIA Level 1 evaluation using identical models and constraints. The comparison highlighted differences in efficiency and task completion between the two agents.
TurboQuant: TurboQuant is an optimization technique that significantly reduces the size of the KV-cache during model inference. It is featured in Atomic Agent to improve efficiency and lower resource demands for extended sessions. The tool complements other caching strategies like stable-prefix caching for cost-effective local operation.
Atomic Agent: Atomic Agent is an open-source, local-first AI operator that runs its full control loop, memory, and task management entirely on the user’s device. It supports web browsing, file operations, document inspection, approved command execution, MCP server connections, skill installation, and Telegram-based instructions while working with both local and cloud models. The agent was introduced with optimizations like stable-prefix caching and TurboQuant for efficient session handling across macOS, Windows, and Linux.

Model Support: Atomic Agent runs multiple open-source models including Qwen, Gemma, and Llama through llama.cpp with features for stable session handling.
Benchmark Comparison: Atomic Agent outperformed Hermes in a controlled GAIA Level 1 evaluation using identical local models, step budgets, and timeouts while publishing full test setups and artifacts for transparency.
Installation and Access: Atomic Agent is available as a one-command open-source install on macOS, Windows, and Linux with additional Telegram-based instruction support.