Fastino launches Pioneer fine-tuning agent

Fastino announced Pioneer, which it calls the first agent for fine-tuning and serving open-source small and large language models. Supported base models include Qwen, Gemma, Llama and GLiNER. Users describe a task in natural language, and the agent handles the training pipeline from there.

Pioneer has two modes. In Agent Mode, users work in a chat interface to generate synthetic data, pick hyperparameters, evaluate models, deploy them and download weights. They can also change datasets, labels or base models between runs. Fastino says users can create a dataset, fine-tune a model and benchmark it against frontier models in as little as 10 minutes. Research Mode runs without supervision and can browse the web. It finds and curates training data, runs experiments in parallel, recovers from failed runs and adjusts based on failure patterns. Fastino says it is meant for jobs that take several hours.

The company also offers what it calls "adaptive inference." Models deployed on Pioneer are retrained over time on production data, including verdicts on outputs from LLM-as-judge and from human reviewers.

Fastino cites GLiNER2, a 205M-parameter encoder that it says matches GPT-4o on named-entity recognition (NER) benchmarks without fine-tuning. The company says GLiNER2 runs on CPU in under 100 milliseconds. Fastino also reports that models fine-tuned on Pioneer reached 99.3% accuracy on intent classification and an F1 score of 0.997 on spam detection. It says those models are more than 500 times smaller than frontier models. The announcement does not name the benchmarks or say how the comparisons were run.

Fastino has since published a research paper on the system, "Pioneer Agent: Continual Improvement of Small Language Models in Production," and summarized its results on its blog. Across eight cold-start benchmarks covering reasoning, math, code generation, summarization and classification, the agent improved on base models by 1.6 to 83.8 points. Fastino also built AdaptFT-Bench, a benchmark of synthetic inference logs with increasing noise that tests the full retraining loop. Pioneer improved or preserved performance in all seven of its scenarios, while a naive retraining baseline lost as much as 43 points. In production-style deployments, intent classification rose from 84.9% to 99.3% and entity F1 from 0.345 to 0.810.

The announcement does not include pricing or availability details.