Stop Guessing About Open Source vs Closed AI Models

The open source vs closed AI models question comes down to ownership. Enterprises should choose open-source models when they want to own their model weights and intellectual property, Jetashree Ravi at Fireworks AI said. One panelist at HumanX 2026 added a warning: open models are being weaponized for attacks against legitimate businesses.

If you build on a proprietary model, Ravi said, you are renting a black box, and the resulting model is never your intellectual property. With open-source models, your fine-tuned model and its weights remain yours.

Open source vs closed AI models: which should your enterprise build on?

Vipul Ved Prakash, Co-founder & CEO at Together AI, recalled that his platform went from serving 10 billion tokens a day in early 2025 to over 5 trillion tokens a day on open-source models. Sujee Maniyam at Nebius reported that the performance gap between open and proprietary models is narrowing.

Maniyam said state-of-the-art models such as Gemini 3.1 and GPT-5 score a 57 on certain scales, while the open-source GLM 5 model scores a 50. The model gap is much narrower than it used to be, he said, and open-source models are very close behind the proprietary ones on pretty much any task.

How close are open models to the frontier?

One panelist argued that open source will always sit slightly below the frontier, about six months behind. Pol Peiffer, Head of Agent Development at Sierra, said open-source models such as Gemma 4 are phenomenal. He said they outperform GPT-4.0 and potentially 4.1, and they can power a pretty good agent.

Why does fine-tuning favor open models?

Ravi described a common scenario. A customer builds an agent that passes 100% of tests in a test environment, then watches it fail to follow required formats consistently in production. That is where the real work begins.

After applying Fireworks' fine-tuning process, Ravi said, the model delivered 12% higher quality and 33% more tool calls. It matched the closed-source model and outperformed it. Maniyam pointed to Cosine, an enterprise coding agent, which trained smaller models on its own data and found they perform better once they are tuned to that data.

How do security and data governance compare?

One panelist recalled an incident where an AI recruitment system went rogue because its default credentials were 123456. The same speaker described another failure, where 700 companies were compromised within 10 days through a third-party dependency.

The panelist argued that routing all enterprise workloads through one company's data center feels extremely dangerous. Open access has its own risk, though. Another speaker in the same session warned that anyone can remove a model's guardrails and run state-scale attacks that could bring enterprises to their knees.

What does it cost to run open models?

Maniyam explained that distillation lets developers take a knowledgeable teacher model and compress it into a smaller student model, and he cited a note-taking app that shrank from 100 billion parameters to 2 billion. Then the infrastructure bill arrives.

Anyone can download an open-source model, Maniyam noted, but running a large one takes hefty infrastructure like GPUs. Because GPU memory is scarce, his team offloads the cache into regular memory or disk so it does not compete with model weights. That caching technique yields speedups of 5 to 10 times. He added that storing a one billion parameter model at full precision uses 4 gigabytes of memory.

How should engineering leaders decide?

Peiffer said a simple turn of conversation in his team's agents can need a dozen or more inference calls, split between in-house fine-tuned models and frontier models. To keep accuracy and speed, they bypass the language model entirely for static knowledge such as legal disclaimers or account balances. The language model produces a template string, and the real value is swapped in at the end.

That constrained problem space pushed Sierra's hallucination evaluations against a knowledge base to near saturation at 99.8%, Peiffer said. When Sierra launched a home search agent for Redfin, the number of listings went up by 2X. His team focuses on outcomes and does not count inference calls, he noted, because Sierra gets paid when its agents do a job well, and not on tokens.

For engineering leaders, the open source vs closed AI models decision rarely has a single winner. Build a system where each model handles the workflow it handles best.

To hear more from the people putting AI into production, see the HumanX 2027 program preview.