Glossary

The terms you’ll meet in the app and these docs, in plain language. Each links to where it matters most.

Auto-evict. Unloading an idle model after a timeout to free memory. Pinned models are exempt. See Memory & loading.

Context length. How many tokens of prompt plus output a model can hold at once. Longer context costs more memory because the KV cache grows with it.

Draft model. A small, fast model that proposes tokens for a larger model to verify, used in speculative decoding. See Performance.

Effective settings. The settings actually applied to a request, after Bedrova merges your global defaults, the model’s profile, and any per-chat overrides. See Model profiles & presets.

Core ML perception. Native Mac image-analysis tasks such as classification, detection, segmentation, depth, OCR, and pose. Perception creates typed artifacts; it is separate from vision chat. See Vision & perception.

Embeddings. Numeric vectors that represent text, images, or supported audio/transcript artifacts so they can be searched by meaning. The backbone of document chat / RAG and multimodal search. See Embeddings.

Fit-check. The pre-load estimate that decides whether a model fits your memory budget, and suggests an alternative when it doesn’t. See What your Mac can run.

GGUF. A model file format used by the llama.cpp ecosystem. Bedrova is MLX-first, but GGUF is supported as a secondary text-only compatibility runtime. When MLX and GGUF are both viable, Bedrova recommends MLX and labels GGUF broad compatibility, usually slower.

JIT (just-in-time) loading. Loading a model on its first request rather than up front. See Memory & loading.

KV cache. The per-token attention state a model keeps during generation. It grows with context length and counts against your memory budget.

MCP (Model Context Protocol). A standard for exposing tools to a model. Bedrova can act as an MCP client for its own chat and API. See Tool calling & MCP.

MLX. Apple’s array/ML framework for Apple Silicon. It’s Bedrova’s primary text/model path, not the only provider pack.

MoE (Mixture of Experts). A model architecture that activates only some “expert” sub-networks per token. Bedrova can expose expert-count control where it’s been shown safe. See Performance.

MTP (multi-token prediction). A speculative-decoding variant that predicts several tokens at once. Supported only where the runtime proves it safe; gated otherwise.

Pin. Mark a model to stay resident so it answers instantly and is never auto-evicted. See Memory & loading.

Quant / quantization. Storing model weights at lower precision (e.g. 4-bit) to cut memory and increase speed, at some quality cost. “Q4”, “Q9”, and “4-bit” are quant levels.

RAG (retrieval-augmented generation). Answering from your own documents by retrieving relevant chunks and adding them to the prompt. See Documents & RAG.

Visual document retrieval. Searching scanned PDFs or page images by visual content, not only OCR text. See Documents & RAG.

Reasoning effort. A mode on reasoning models that trades latency for deeper deliberation (for example non-think / high / max). See Chat & reasoning.

Speculative decoding. Using a draft model to propose tokens that the main model verifies in bulk, increasing throughput. See Performance.

Structured output. Forcing the model’s response to match a JSON schema. See Structured output.

TTFT (time to first token). How long after you send a request before the first token streams back. Reported live in the token inspector.

TTL (time to live). The idle timeout after which a just-in-time model auto-evicts.

Unified memory. The single memory pool Apple Silicon shares between CPU and GPU. It’s the budget every loaded model draws from. See Concepts.

Wired memory. Memory the system holds resident and won’t page out. Model weights on the GPU are effectively wired; it shows up in the hardware monitor.