Freesolo

The problem Freesolo is attacking is real and boringly practical: for a long tail of tasks the intelligence already exists in a small model, but serving it at acceptable latency and cost does not. The vendor's own framing on the homepage is that engineers have to choose between model size and quality, that post-training closes the gap, and that the activation energy to train is currently far too high. Freesolo's answer is to make the training loop agent-shaped: you point Claude Code, Cursor, or Codex at the Flash package and let the agent iterate. Strengths that show up concretely: fixed-price quoting per run instead of metered tokens, which removes the scariest part of a training bill; full ownership of exported weights in standard formats; support for all three common post-training recipes (SFT, GRPO, OPD) plus multimodal SFT/GRPO/OPD for vision inputs; custom kernels tuned for H100 and A100; CUDA graphs that cut decode latency by 7-10x for 9B and 27B models; multi-GPU training via flash train --gpu.count; a Flash CLI that now runs on Windows; live per-step metrics and auto-refreshing logs; and reproducible runs with pinned configs and seeds. Data is encrypted in transit and at rest and is never used for other training. The company also publishes its own agent work, e.g. a browser agent trained with DSPy-style recursive prompt optimization that solved a 30-step challenge in 265 seconds at $0.06 per run with GPT-OSS 120B matching frontier performance. Weaknesses and honest friction: it is agent-native and CLI-shaped, so a GUI-first user will feel out of place. You need to bring example data and, for GRPO, a reward function; the platform will not invent your task definition. It is a fine-tuning and post-training tool, not a pre-training service, so you cannot use it to train a foundation model from scratch. The breadth of high-quality third-party documentation is thinner than a large incumbent's, and the integration surface is narrow: Claude Code, Cursor, Codex, and Fireworks appear in the vendor's own materials. Benchmarks come mostly from Freesolo's own blog, so treat vendor-reported speed and cost numbers as directional until you run your own task. Where it fits: teams with proprietary labeled data who want to cut per-query cost on a narrow task (support-ticket classification, query routing, extractive QA on internal docs, autocomplete, vision classification), and teams who want to own the weights so they can move them onto their own serving stack later. Where it does not: anyone whose problem needs frontier-level open-ended reasoning, anyone who cannot define a reward signal or prepare examples, and anyone wanting a point-and-click tool.