What it means
An open weight large language model (LLM) is a model whose trained parameters are released publicly, so you can download, run, fine-tune, or self-host it instead of only reaching it through a vendor's API. The weights are open, but the training data and code often are not, which separates open weight from fully open source. Llama, Mistral, and DeepSeek are common examples.
Why it matters
Open weight models give you direct control over LLM capability, but the same hallucination and citation risks still apply. When you can self-host a model, you are not locked into one vendor's pricing, rate limits, or data policy, and you can fine-tune on your own data.
The reliability problems Kevin flags for any LLM do not disappear: Models make things up convincingly, and different models pull from different citation sources, so output quality varies.
Say a mid-market B2B SaaS company self-hosts an open weight model to summarize support tickets. It cuts per-query cost from $0.02 on a proprietary API to near zero at scale, but a 5% hallucination rate means answers still need human review before they reach customers.
How to use this knowledge
Decide per use case whether control (self-hosting, fine-tuning, data privacy) outweighs the quality edge of a frontier API model. Benchmark an open weight model on your actual tasks, not public leaderboards, and measure the hallucination rate you can tolerate. Keep a human in the loop for anything customer-facing. Track which models cite which sources, since that shapes whether your content gets surfaced.
Growth Memo guidance
The biggest risk of using generative AI and, at the same time, the biggest factor holding even wider adoption back is hallucination. LLMs can make things up in a very convincing way. — #6 - The state of generative AI for SEO (PT.2)
Important to note is that LLMs rely on different citation sources to varying degrees. — AI Halftime Report H1 2025
If content is the new oil, social networks are oil rigs. — Labeled
Large language model (LLM) — the broader model class that open weight and closed models both belong to.
Hallucination — the reliability risk that persists whether a model is open weight or proprietary.
Fine-tuning — the main reason to pick an open weight model, since you can retrain it on your own data.
Training data — open weights ship the parameters but not the data the model learned from.