When to choose open source vs closed AI models for enterprise production
Fine-tuning, IP ownership and how fast open models are closing the gap.
Enterprises weighing open source vs closed AI models have more real-world data to go on than they did a year ago. Tom Mackenzie at Bloomberg cited statistics showing that between 10% and 40% of Fortune 500 companies are using a multi-model mix or increasingly using open weights.
Jetashree Ravi, tech lead manager at Fireworks AI, made the case for open source on two fronts. Fine-tuned open models can outperform closed models on complex agent tasks, and when you fine-tune a closed-source model, the result is not your intellectual property.
Why do closed models fail in production?
Ravi described customers whose agents passed “100%” of the tests written in the test environment, then failed once they were rolled out. In production, the agents didn't follow required formats consistently and picked the wrong tools. She described closed-source models as a complete black box.
How does fine-tuning change the open source vs closed AI models equation?
Sujee Maniyam, developer advocate at Nebius, recalled that the enterprise coding agent Cosine trained smaller models on its own data instead of relying solely on a 300 billion parameter model, and that made them perform better.
Ravi shared a comparison from Fireworks. A closed proprietary model achieved a reward score of 0.6 with an average of 3.74 tool calls. The Kimi K2 model, fine-tuned on Fireworks, scored 0.82 with an average of five tool calls.
Where does open source stand against the frontier?
Maniyam said the top proprietary models score around 57 on Artificial Analysis composite benchmarks, while the leading open-source model, GLM 5, scores 50. That gap is much narrower than it used to be.
Thomas Wolf at Hugging Face said the latest estimate puts open-source models roughly six months behind the closed-source frontier. He argued that it is much harder to push the frontier than to catch up to it, which keeps open source competitive. Wolf recalled that “DeepSeek was the first time actually one open source model was slightly above on some metrics.”
How do enterprises manage IP and sovereignty?
Umesh Sachdev at Uniphore said sovereign AI in the enterprise is about optionality, flexibility and control over the AI stack.
Wolf warned against centralizing AI workloads in one company, describing a future where “all of our workload transits through just one data center of one company. To me, it feels extremely dangerous, actually.”
What is the adoption rate at scale?
Vipul Ved Prakash of Together AI described open-source adoption as massive: “we went from altogether serving 10 billion tokens a day last year, this is early 2025, now serving over 5 trillion tokens a day.”
Prakash argued that Chinese open-source models lead the ecosystem because a cultural indifference to the open or closed distinction drives a game theory of building better models.
For enterprises choosing between open source vs closed AI models, Ravi's argument comes down to ownership. Fine-tune an open-source model and the weights and the IP are yours, and in her comparison the fine-tuned open model beat the closed one.
To hear more from the people putting AI into production, see the HumanX 2027 program preview.
Frequently asked questions
When does fine-tuning an open-source model outperform closed models for enterprise use cases?
Jetashree Ravi at Fireworks AI argued that fine-tuning an open-source model on specific workflows improves its reward score and tool-calling capacity beyond proprietary baselines. She reported that this approach delivered higher quality and more tool calls than the closed-source model.