While frontier AI headlines go to OpenAI and Anthropic, a parallel story has been developing in open source. Mistral's models and the Llama ecosystem have produced models that are genuinely useful in production.
Mistral 7B and what small models can do
Mistral 7B, released in September 2023, was a demonstration that a 7 billion parameter model trained carefully could outperform Llama 2 13B on most benchmarks. The mixture-of-experts architecture in Mixtral 8x7B (released December 2023) took this further: routing tokens to specialised expert networks produces a model with 46.7B total parameters but only 12.9B active per token, giving it the throughput of a 12.9B model with the capability of a much larger one. Mixtral 8x7B outperforms Llama 2 70B on most benchmarks while being three times faster at inference.
The Llama ecosystem
Llama 2, released by Meta in July 2023, legitimised enterprise use of open source LLMs. The Llama licence permits commercial use with attribution and without a royalty, for organisations below a certain monthly active user threshold. The fine-tuning ecosystem that has grown around Llama 2 is substantial. Tools like Axolotl, LitGPT, and Unsloth make fine-tuning accessible. Quantisation tools like llama.cpp let you run 70B models on consumer hardware. The gap between open source capability and proprietary API capability is narrowing.
What organisations are doing with open models
The most common enterprise pattern with open source LLMs is running a smaller model on-premises for classification, extraction, and routing tasks, while calling a frontier model API for generation and complex reasoning. The economics work: a Mistral 7B instance on a single A10G GPU costs fractions of a cent per request at scale. The privacy case is also clear: data never leaves your network. Open source fills the gap between rule-based automation and the power of GPT-4, at a cost and control profile that suits many compliance environments.
The dynamic with proprietary models
Proprietary model providers are aware of the open source gap-closing. OpenAI and Anthropic are investing heavily in distillation and efficiency research. The frontier is moving faster than open source can match, but the frontier is also not where most production AI workloads live. The majority of tasks that enterprise software needs to handle, classification, extraction, summarisation, routing, structured data generation, can be done effectively by well-tuned open source models at a fraction of the cost of frontier API calls.