The production platform for open-weight AI inference

The production platform for open-weight AI inference Webflow Analyze/Optimize tracking bridge -->

💰 Announcing our Series C. Intelligence should be abundant, not expensive →

🤝 Together AI & Y Combinator announce partnership to deliver the first dedicated YC GPU cluster →

⚡ On-demand B200s now available on Together GPU Clusters →

🚀 Now serving MiniMax-M3 for efficient inference →

All blog posts

Inference

Published 7/23/2026

The production platform for open-weight AI inference

Deploy in minutes. Roll out safely. Scale to your SLOs.

Authors

Nikitha Suryadevara, Ted Cui, Will Van Eaton, Charles Zedlewski

Table of contents

40+ Models Chosen for Production...40+ Models Chosen for Production...40+ Models Chosen for Production...

Links in this article

Get started Talk to our team Read the docs Sign up for the webinar

Summary

We're releasing a significant update to our inference platform, giving you complete control over performance, cost, and quality without building your own stack. Models go live in minutes and every deployment is production-grade from the start: run multiple deployments behind one stable endpoint, ship changes safely with canary, blue-green, and rolling updates that auto-roll-back on your thresholds, test on real traffic with A/B and shadow testing, and autoscale across one region or many. We're also announcing a closed beta for custom training, including full-weight and LoRA reinforcement learning and supervised fine-tuning, with checkpoints you can deploy straight to production. Request access to the custom training beta. Launch your first endpoint today. Join our webinar on August 6 for a deep dive.

Open-weight models now match closed models on quality, run at a fraction of the cost, and are fully customizable for your task. That combination has made them the foundation for teams building serious AI products and agents. But the strategic motivation for adopting open-weight models is still the ability to exercise control. Unlike their closed alternatives, open-weight models give teams control over performance, quality, and functionality. Many commercial AI applications have adopted a majority open-weight endpoint strategy so they can maintain complete control over their user experience. Many enterprises use this control to incorporate their valuable proprietary IP into models without risking exposure to third parties. While open-weight model users want maximum control over performance, functionality, and quality, few teams want to test their ability to correctly align a matrix of decisions across quantization levels, parallelism schemes, engine parameters, and draft model architectures—to name just a few—on a weekly basis. Meanwhile, frontier techniques for achieving the best quality and performance are advancing just as quickly. In a rapidly evolving AI landscape, no one has the time or desire to stop and read an almanac’s worth of trivia about inference internals or a mountain of research papers—not even the teams running the world’s most successful models and agents. Together Dedicated Model Inference applies the lessons we’ve accumulated from serving more than 400 trillion tokens per month to deliver an inference platform that: Gives users complete control over their models, performance, cost, quality, and functionality Prevents lost time, wasted spend, and unintended downtime from building each inference stack and endpoint management layer from scratch Continually incorporates the latest vetted research to achieve frontier levels of performance, quality, and efficiency

This update brings together advances from our research, model optimization, and platform engineering teams in a single service designed to give companies an easier, more reliable, and more optimal way to run open-weight, licensed closed-weight, and fine-tuned models in production. Delivering that combination of control and simplicity requires treating inference as more than hosting a model behind an API. The precision you run, the hardware you deploy on, the serving configuration you use, and the way you scale and route traffic all have a meaningful impact on performance, reliability, and cost. We built this platform so that moving from experimentation to production does not require changing platforms or rebuilding your deployment. Production readiness is built into the foundation, even when you are still testing and iterating. That means you can roll out new versions safely, test changes against real traffic, shadow production requests, route traffic across deployments, scale with demand, and roll back if something does not work as expected. Max Lu at Decagon described the shift this way: Together already got our latency where it needed to be for voice. What their new inference platform changes is how we ship the next model version: we can canary a new fine-tuned model on a small percentage of live traffic and automatically roll back if key metrics regress. We will go from ‘fast inference’ to ‘fast inference we can safely iterate on every week.’ That is ultimately what we wanted to unlock: not just fast, efficient inference, but the ability to continuously improve what you are serving without introducing unnecessary risk. You can start with the paths Together has already optimized, then take on more control as your workload becomes more specialized. That principle is reflected across the platform—from how you bring in and configure a model to how you scale it, test changes, observe performance, and move new versions into production.

Bring the model you want to serve Deploy models from Together’s model platform, or bring your own open-weight or fine-tuned model. You can upload full model weights or adapters from Hugging Face, S3, or your local machine. The platform is designed to support the full lifecycle of a model, from an early fine-tune through the versions that eventually serve production traffic. Choose how the model runs The model itself is only one part of the deployment. Hardware type, quantization, tensor parallelism, speculative decoding, and the serving engine all affect the performance and economics of the workload. You should not have to become an expert in every one of those choices to get strong results. Together deployment profiles package configurations our research and engineering teams have already tested and optimized, so you can start with a proven setup rather than assembling one from scratch. When you need more control, those choices are...