The AI terms that
actually matter.
Clear, technical definitions of 148 key concepts — written by AI engineers, not marketers.
Mixture of Depths (MoD)
Mixture of Depths is a transformer routing technique where a learned per-layer router selects only a fixed fraction of tokens to pass through that layer's full attention and MLP block; the rest skip the block entirely via a residual connection. Because the fraction (capacity) is fixed ahead of time, total compute stays static and predictable, unlike Mixture-of-Experts routing.
Read definitionMixture of Experts (MoE)
Mixture of Experts is a neural network architecture where a learned routing mechanism activates only a small subset of specialised sub-networks (experts) for each input token — delivering the capacity of a much larger model at a fraction of the per-token compute cost.
Read definitionModel Context Protocol (MCP)
Model Context Protocol (MCP) is an open standard introduced by Anthropic that defines a universal interface for connecting AI models to external data sources, tools, and services — eliminating the need for bespoke integrations for every AI-to-system connection.
Read definitionModel Merging
Model merging combines the weights of two or more independently trained or fine-tuned neural networks into a single model, without backpropagation or extra training, producing a model that inherits multiple source models' capabilities at zero added inference cost.
Read definitionModel Retrieval and Knowledge Graphs
Model retrieval and knowledge graphs refers to grounding a language model's answers in a structured graph of entities and relationships rather than (or alongside) unstructured text chunks, most commonly through GraphRAG-style pipelines. It trades the cost of building and maintaining a graph for multi-hop reasoning and query-focused summarization that flat vector retrieval cannot reliably do.
Read definitionMoshi
Moshi is a full-duplex, real-time spoken dialogue AI developed by Kyutai that can simultaneously listen and speak — enabling natural, interruption-aware conversations with theoretical latency as low as 160ms.
Read definitionMulti Agent System
A Multi-Agent System (MAS) is an AI architecture where multiple autonomous, specialized agents interact, collaborate, or compete to solve complex workflows that are too difficult or broad for a single, monolithic AI model.
Read definitionMulti-Head Latent Attention (MLA)
Multi-Head Latent Attention (MLA) is the attention variant introduced by DeepSeek-V2 that compresses each token's key and value vectors into a single low-rank latent vector, shrinking the KV cache far more aggressively than Grouped Query Attention while matching or beating full multi-head attention's output quality.
Read definitionNatively Multimodal Model
A natively multimodal model is trained from scratch on multiple modalities at once, text, images, audio, video, rather than bolting a frozen vision or audio encoder onto a finished language model. The modalities share one token stream and one set of weights, which is what 'native' refers to.
Read definitionNVIDIA Cosmos 3 Edge
NVIDIA Cosmos 3 Edge is a 4-billion-parameter open world foundation model, the smallest tier of the Cosmos 3 family, that reasons over vision and generates robot actions directly on Jetson-class edge hardware without a cloud round-trip.
Read definitionNVIDIA Parakeet TDT
NVIDIA's Parakeet TDT 1.1B is the fastest open-source ASR model available, achieving an RTFx near 2,000× real-time — processing audio 6.5× faster than Canary-Qwen at the cost of some accuracy.
Read definitionOKF (Open Knowledge Format)
OKF is an open specification from Google Cloud, published June 2026 under Apache-2.0, for representing an organization's curated knowledge as a directory of plain Markdown files with YAML frontmatter and explicit links between concepts. It is meant to be authored by humans and read (and updated) directly by AI agents, keeping relationships explicit rather than re-deriving them at query time the way RAG does.
Read definitionStay ahead of the curve
Weekly newsletter on agentic AI, LLMs, and what we're building at Superteams — straight to your inbox.
Ready to ship AI in production?
We deploy fractional AI teams that deliver production-grade systems in 30–90 days. No fluff, no obligation.