Small Models, Big Wins: Fine-Tuning Open Weights for the Enterprise

For roughly eighteen months the answer to "which model should we use?" has been "whichever frontier lab shipped last week". That's finally changing. Between Meta's Llama 3.3 release and the wave of highly-capable Mistral and Qwen checkpoints, small, self-hosted models are now genuinely competitive for a widening set of enterprise tasks.

The interesting question isn't "can I run an LLM myself" — it's "when should I".

The four situations where we recommend self-hosting

• Data residency is a hard requirement (regulated industries, EU public sector, healthcare).
• The workload is narrow, high-volume and repetitive — think document classification, entity extraction, structured summarisation.
• Latency budgets are tight enough that a round-trip to a frontier API is the bottleneck.
• Total token volume has grown past the point where per-token pricing beats a reserved GPU.

Our default fine-tuning recipe

We start with Llama 3.3 8B or Mistral 7B, run a LoRA fine-tune on 2–10k task-specific examples, and quantise to 4-bit for serving. On the classification and extraction tasks we see most often, this consistently reaches 95%+ of GPT-5 mini quality at a fraction of the ongoing cost.

The trick is not the training run. It's the data. We spend most of the project turning messy real-world traces into a clean, labelled dataset — and the model does the rest.

The hybrid pattern nobody talks about

Our favourite architecture in 2025 was a small self-hosted model doing 90% of the volume, escalating only edge cases to a frontier API. It's cheaper, faster, more private and, unfairly, often more accurate — because the small model is fine-tuned exactly on your task.