Nebius and Eigen AI partner to accelerate frontier open-source AI inference

Nebius and Eigen AI are partnering to bring faster, optimized open-source AI models to Token Factory, Nebius’s production-grade managed inference platform.

As part of the collaboration, Nebius and Eigen AI are co-developing optimized versions of leading open-source models — including DeepSeek, GLM, GPT-OSS, Kimi, Llama, MiniMax and Qwen, — and integrating them into Token Factory. Eigen brings deep expertise in model optimization and serving systems, while Token Factory provides autoscaling inference and built-in fine-tuning tools.

Developers can access these models through an API on a per-token basis, or run them as managed solutions for production workloads.

Running open models in productionRunning open models in production

More organizations are moving to open-source AI models. They offer lower cost compared to proprietary APIs and allow teams to customize models for their own data, workflows and infrastructure.

At the same time, many of the newest open models — including Mixture-of-Experts (MoE) architectures, Linear Attention variants and reasoning models — are harder to run efficiently at scale. Getting strong performance requires optimized inference runtimes, smart GPU scheduling and infrastructure designed for large models.

Without a production platform, teams typically have to run these models themselves — often building custom infrastructure around frameworks like vLLM, Ray or Kubernetes, managing GPU clusters, tuning inference performance and maintaining scaling and reliability on their own. This adds significant engineering overhead and makes it difficult to move quickly from experimentation to production.

Token Factory is designed to close that gap. It provides a production platform for running, improving and operating open-source models.

Key capabilities include:

• Autoscaling inference endpoints that adjust capacity automatically as traffic changes;

• Dedicated model endpoints with guaranteed performance isolation and service levels;

• Integrated post-training pipelines for LoRA fine-tuning and distillation;

• Draft model training for speculative decoding to improve inference efficiency;

• Instant promotion of tuned models into production endpoints for fast integration;

• Enterprise governance tools, including team workspaces, SSO and access controls.

Together, these capabilities allow AI developers to adapt open models to their own data and run them in production without managing infrastructure themselves.

Optimized open models and serving from Eigen AIOptimized open models and serving from Eigen AI

Eigen AI specializes in making frontier open-source models fast and efficient in production through deep full-stack optimization. At the model layer, Eigen improves efficiency with advanced post-training quantization, quantization-aware training, KV-cache optimization and multi-granular sparsity techniques that reduce compute and memory costs while maintaining strong model quality.

At the systems layer, Eigen improves how these models run in production. Their work includes speculative decoding, custom CUDA and Triton kernels, parallel execution, continuous batching and graph-level runtime optimizations.

In practice, these optimizations help models generate tokens faster, use GPUs more efficiently and reduce the cost of serving large models at scale. This is especially important for modern Mixture-of-Experts and reasoning models, where routing, scheduling and memory efficiency often determine whether a model can run reliably and economically in production.

By bringing these optimized implementations into Nebius Token Factory, Nebius and Eigen AI are making it easier for developers to use frontier open models with high speed, reliability and production readiness, without having to build and maintain the optimization stack themselves.

Leading output speeds across popular open modelsLeading output speeds across popular open models

Eigen’s optimized models have demonstrated leading performance in benchmarks tracked by Artificial Analysis.

For example, Eigen currently holds the #1 output speed, for multiple widely used models, reaching up to 911 output tokens per second.

Eigen Achieves #1 Output Speed Across Leading Open ModelsEigen Achieves #1 Output Speed Across Leading Open Models

(Artificial Analysis benchmarks, as of March 13, 2026)