The Frontier is Open

The Frontier is Open

⚡️ FlashAttention-4: up to 1.3× faster than cuDNN on NVIDIA Blackwell →

Introducing Together AI's new look →

🔎 ATLAS: runtime-learning accelerators delivering up to 4x faster LLM inference →

⚡ Together GPU Clusters: self-service NVIDIA GPUs, now generally available →

📦 Batch Inference API: Process billions of tokens at 50% lower cost for most models →

🪛 Fine-Tuning Platform Upgrades: Larger Models, Longer Contexts →

All blog posts

Company

Published 6/9/2025

The Frontier is Open

Authors

Charles Zedlewski

Table of contents

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

Links in this article

Charles' LinkedIn

Peter Thiel once said that good startups are built around secrets: ideas that only a few people see at first but that becomes obvious later. There’s no secret that AI models are the foundation of our next application platform; it’s arguably the most seismic platform shift since the internet roughly 30 years ago. What is still a secret to most in 2025, is this new AI application platform is quickly becoming a ubiquitous open source commodity. Openness and open source has been the inevitable direction of nearly every new platform dating back to the internet browser. From browsers to operating systems, to databases to containers, users have converged on open source as the “default” path for usage and innovation. This shift from proprietary to open source is progressing even faster this time with AI. AI models are built on open research and trained on public data. It’s odd that many of the best known AI models today are proprietary derivations of such open ingredients. Open source platforms are commonly known to offer greater flexibility and lower cost. These qualities matter more for AI applications than they have in past platforms. The cost advantages of open source in AI are considerable both because: AI models have similar interfaces but are subtly incompatible in their respective strengths and biases - open source means you can remain invested in a model but still have portability and accountability from infrastructure providers The compute to power AI is still quite expensive and open source has been in the lead in innovating on the efficient use of this precious resource

The flexibility of open weights and open source licenses means you can: Adapt models to your use case in ways you can only do with open weights: for example quantize them to fit your quality/performance requirements Own the final artifact so you can run your AI where it needs to be run whether that be a cloud, a device, a robot or an airgapped datacenter Distill within and across model families without fear of violating proprietary license terms

Not only is open source AI more flexible and less costly than proprietary, it’s also increasingly where we find the frontier for AI. Because AI is a relatively new and immature technology, we’re understandably compelled to run towards the frontier to reduce our risk of failure. Proprietary AI developers want to believe the frontier can only progress from closed labs working around a magic cauldron sitting on a pyre of GPU’s. But building a future on proprietary AI is betting that success is best secured by the sum of the smart people working at just one lab. As Bill Joy once said: “ no matter who you are, most of the smartest people work for someone else .” Today open research is producing open models that occupy different points on the current AI frontier. Open source models are claiming more “firsts”: Llama 3 was a breakthrough in AI affordability. It was the first source of quality, low cost intelligence and brought down industry API pricing by 80%. Deepseek R1 was a breakthrough in the use mixture of experts architecture to make very large gains in pretraining & inference efficiency Qwen3 was the first ever foundation model designed for “hybrid reasoning”

And open source models already hold a number of the superlatives in the ecosystem of open and closed models. Today: The largest range of model sizes is in open source, enabling developers to find the best fit for their use case The highest quality instruct model is open source, meaning the best go-to model for fast, consistent intelligence for agent building is open source The most parameter efficient SOTA quality coding model is open source - making it possible to have powerful AI coding assistance local to a laptop The best source for new model distillation is a mixture of open source models

Proprietary models continue to occupy the frontier in several dimensions such as reasoning and multimodality, but these gaps have been shrinking over time. It would be naive to think open source won’t someday have SOTA frontier achievements in these areas as well. Open source is the way forward for AI as a platform — but for developers to benefit from this trend, open source AI needs a new kind of cloud. This is the reason why Together AI was founded by Vipul, Ce, Tri, Percy and Chris, because: Open out-innovates thanks to community diversity, but diverse also means variable. There are nearly 2 million open source models on Huggingface but they each have unique configurations, dependencies and infrastructure needs. Together AI is a cloud platform that normalizes this variability so we can have production uses of open source AI. Open-source models are more adaptable to your use case but the skill expectations of that adaption can be excessive. Together AI is a cloud platform that gives powerful, usable abstractions so engineers—not just researchers—can fit AI models to their use case. Open source models are architected assuming AI-native infrastructure which is evolving almost as fast as the models. Together AI is a cloud platform that’s built to quickly adopt the newest AI native compute, networking & storage architectures to absorb the next wave of open source innovation.

Together AI was built to serve as the bridge between developers, researchers and the open source AI frontier. The signs of the founders’ ambition are everywhere. Together AI has the best roster of researchers of any open AI platform company today because it takes researchers to translate frontier research into a production platform whose results can match or exceed the benchmarks in papers. Together AI has built the most comprehensive range of pretraining & posttraining services so users can start at any place in the AI development…