Freesolo Flash

Full-stack training platforms for SLMs feel like the right direction as more people realize they don't need a massive model for most tasks. Does it handle the data prep/cleaning side too, or is that still on the user before they get to training?

Hey. Upfront pricing quoted before the run starts is the part I'd stress test, since that only works if the cost estimate is actually accurate to what the training ends up needing. RL runs in particular are notoriously unpredictable in how many steps or rollouts it takes to converge, especially on a narrow task where the reward signal might be noisy early on. If a GRPO run needs meaningfully more steps than estimated to actually reach a usable policy, does Freesolo eat that overage to honor the quoted price, or does the agent get cut off at the budget with a model that never really finished training.

Also curious how the environment hub interacts with that pricing model. If someone builds a custom environment through your SDK that behaves in some unexpected way, say a reward function that's easy to game or a slow environment step, does that variability get priced into the upfront quote too, or is upfront pricing really only reliable for the standard environments you already understand well.

Small language models are having a moment for good reason, the economics of running a frontier model for every agent step do not hold up at scale. Lowering the barrier to training your own SLM is a useful place to build. One angle for positioning, show the total cost delta of an SLM-first agent versus an all-LLM one. That comparison sells itself to anyone watching their inference bill.

Not every AI interaction is best served with a large frontier model. There is a long tail of trillion-token use cases, from tagging to search, best served by a sub-10b-parameter model that runs in milliseconds and costs many orders of magnitude less than the frontier. However, engineers historically had to choose between model size and quality; as models got smaller, performance, adherence, and recall dropped linearly. Post-training on production data closes that gap. We built Freesolo Flash to make training loops like SFT and RL easy and end-to-end completable through your coding agent.

We accomplish this in a few different ways:

- Upfront pricing: instead of billing by GPU hours or tokens spent while training, we quote the cost of the entire run upfront, so your agent can accurately tweak the dataset, model size, and algorithms it uses while staying in your budget before it starts the run.

- Our GPU infrastructure is optimized to make your specific run as in-expensive and fast as possible. This optimization means training with flash is 8x less expensive for SFT and 5.5x less expensive for GRPO (RL) when compared to Tinker.

- Environment hub: Our custom environment SDK allows you to build environments in a modular way and perfectly integrates with our asynchronous training framework.

Flash is built out of our own frustrations with current managed post-training solutions, especially for SLMs. We believe that unlocking frontier capability for a narrow task into a small model will prove to be the best improvement for all agentic product ux. Flash is our first step towards solving this. Just grab a Freesolo API Key, point your agent at the training package, and watch it push your lightweight model beyond the frontier.