Author: Aethir

On June 3, 2026, Aethir officially opens Aethir Mesh to all developers and AI agent operators—this is Aethir's self-developed open-source large model API platform, where all inference tasks run directly on Aethir's decentralized GPU infrastructure.

Direct computing supply, not a routing intermediary

To understand the positioning of Aethir Mesh, it is essential to grasp its fundamental differences from existing solutions.

Mainstream open-source model API platforms, represented by OpenRouter, are essentially an aggregation routing layer—forwarding user requests to third-party inference providers' servers such as OpenAI, Anthropic, and Google, with the platform taking a resale margin. User data inevitably flows through external servers in this process.

The architecture of Aethir Mesh is entirely different: open-source models run directly on Aethir's own decentralized GPU infrastructure—over 430,000 GPU containers, covering 94 countries, and more than 200 nodes. Aethir is the layer that performs computation, rather than just packaging others' APIs for billing. This means no resale margin, no third-party data routing, and an enterprise-level service standard with GPU utilization maintained at over 95%.

11 models, single endpoint

Aethir Mesh currently features a total of 11 models from five major open-source providers, with the core lineup including:

• DeepSeek V4: 1.6 trillion parameter MoE architecture, activating only a small portion at each inference step to achieve performance close to cutting-edge closed-source models at significantly lower inference costs, suitable for most general tasks of agents;

• Kimi K2.6: Produced by Moonshot AI, 1 trillion parameters MoE, with 32 billion active parameters per step, natively supports large-scale sub-agent collaborative orchestration, designed for multi-agent pipelines;

• GLM-5.1: Produced by Z.AI, under MIT open-source license, MoE architecture, with 40 billion active parameters per step, ultra-long context window designed for long-range tasks, supports continuous operation up to 8 hours and thousands of tool calls;

• MiniMax-M2.5: Released in February 2026, targeting programming, agent tool usage, web search, and complex office task scenarios, with an inference speed of 206.7 tokens/second;

• Qwen3.6-27B: Produced by Alibaba, 27 billion parameters, under Apache 2.0 license, natively 262K Token context (scalable to millions), surpassing the previous flagship Qwen3.5-397B across major coding benchmarks.

All models are accessible via a unified API endpoint, allowing developers to manage a single Key, single billing, and single dashboard, without the need to maintain multiple authentication processes across platforms.

Full-stack integration with Aethir Claw

For users of the Aethir Claw agent platform, Aethir Mesh is the final piece in their full-stack closed loop. After writing the Mesh API Key into the agent configuration on the Aethir Claw dashboard, all model inference requests for that agent will automatically route through Aethir Mesh and return results within the platform—VPS instances, skill execution layer, and model inference layer share the same infrastructure, unified billing, and unified monitoring, with inference data never leaving the Aethir ecosystem.

For independent developers and enterprises not using Aethir Claw, Aethir Mesh can also serve as a standalone inference endpoint, suitable for Agentic AI application development, DePIN projects, and enterprise-level AI workflows.

Continuously expanding model catalog

The model catalog of Aethir Mesh is designed to dynamically expand with the open-source ecosystem. As high-performance open-source models are released and achieve production stability, Mesh will continuously incorporate new models, allowing users to access the latest capabilities without switching API providers or renegotiating access terms.

Technical documentation , Get started now