Companies choosing between open-source models (Llama, Mistral, Qwen) and closed-source APIs (GPT-4, Claude, Gemini) often ask which hallucinates less. The answer: openness is not the variable. Model size, alignment quality, and task specificity determine hallucination rates. Both share the same architectural gap.
Closed models hold most of the top places. On Vectara's leaderboard for grounded summarisation (22 September 2026), GPT-5.4 nano scores 3.1% and Gemini 2.5 Flash-Lite 3.3%. They benefit from massive compute budgets, proprietary training data, and extensive tuning.
Open-weight models are close behind. On the same leaderboard, Microsoft's open-weight Phi-4 scores 3.7%, within one percentage point of the best closed models listed. On specific tasks, open models match or beat their closed counterparts.
Smaller open-source models (7B-13B parameter range) still hallucinate noticeably more. These are the models most commonly deployed locally or on-premise. Their smaller size means less training data coverage, less nuanced pattern matching, and more frequent knowledge gaps to fill with fabrication.
Model size. Larger models know more, but knowing more is not the same as knowing what you don't know. The AA-Omniscience authors found that "greater scale does not translate directly into higher knowledge reliability." This is not about openness. But even the largest models still hallucinate.
Alignment quality. RLHF and other alignment techniques reduce hallucination - but can also create new forms of it (sycophancy, over-confident gap-filling). Closed-source providers typically invest more in alignment. But some open-source models (Llama 3.1, Qwen 2.5) now have competitive alignment.
Task specificity. A model fine-tuned for a specific domain often hallucinates less in that domain than a larger general-purpose model. An open-source model fine-tuned on medical data may outperform GPT-4 on medical questions - but may hallucinate more on unrelated topics.
Training data quality. The quality and recency of training data matters more than the size of the model. A smaller model trained on high-quality, recent data can outperform a larger model trained on older, noisier data.
The model's license does not determine whether it hallucinates. The architecture does. Open and closed models share the same architecture: generate, then deliver. Neither generates, then verifies.
Open-source models have one advantage that closed-source models do not: transparency. You can inspect the model's weights, monitor its internal representations, and build custom verification pipelines around it.
Logprob access. Open-source models give you full access to token logprobs. You can build uncertainty detectors that flag low-confidence outputs. Closed-source APIs may limit or omit logprob access.
Custom grounding pipelines. With an open-source model, you can build verification layers directly into the inference pipeline. You control the architecture end to end. With a closed-source API, you are limited to the provider's exposed interface.
But the fundamental gap remains. Whether open or closed, the model still generates without verifying. Logprob access helps detect some hallucinations but not all. The fix is not model selection - it is adding a verification layer that checks output against reality. Check works with any model, open or closed, because verification is model-agnostic.
14 days free. Get your GOL API Key at the console.