LLM
Large Language Model. A neural network trained on vast amounts of text to predict the next word in a sequence. Modern LLMs (GPT, Claude, Llama, Qwen, DeepSeek) generate human-quality text and are the foundation of most modern AI products.
Also known as: Large Language Model, language model
LLMs are the type of AI model that almost everyone now interacts with daily. They’re trained on massive text corpora (large fractions of the internet, books, code, papers) using a simple objective: predict the next token (roughly, the next word) in a sequence. After enough training, the model develops an internal representation of language so rich that it can produce coherent essays, answer questions, write code, summarise documents, and have conversations. None of this was hand-programmed; it emerged from learning the statistics of human-generated text.
The major LLM families today are roughly OpenAI’s GPT series (GPT-4, GPT-5, mostly closed-weight), Anthropic’s Claude series (Claude 3.5, Opus 4.6, closed-weight), Google’s Gemini, Meta’s Llama (open-weight), Alibaba’s Qwen (open-weight), DeepSeek (open-weight), and Mistral (mixed). The performance gap between closed-weight frontier models and the best open-weight models has narrowed dramatically: as of 2026, top open-weight models match or beat older versions of the closed leaders on most benchmarks while being free to download and run.
The critical thing to understand about LLMs is that they have no real-time knowledge, no persistent memory between conversations (unless explicitly added by the application), and no understanding in any deep sense. They’re pattern-matchers that have absorbed enormous amounts of text and can produce text that fits the patterns of the prompt. This doesn’t mean they’re useless or fake; pattern-matching at this scale produces genuinely useful outputs for many tasks. But it does mean their failure modes (confidently wrong “hallucinations”, inability to reason about novel situations, manipulation by adversarial prompts) are different from human failure modes and need to be understood separately.
In DeAI, LLMs are the primary workload for decentralised inference projects. Venice runs dozens of open-weight LLMs through its private inference gateway. Morpheus’s marketplace lets users request inference from any registered model. Bittensor subnets compete on producing better LLM outputs through their incentive mechanism. Most of OYM’s project reviews of inference projects are ultimately about how well the project serves LLM workloads with specific privacy, sovereignty, or cost properties.