LLM Glossary
The terms you'll meet on every leaderboard — explained in plain language and linked to the live data behind them.
Podium Score
LLMPodium’s 0–100 composite ranking score: 0.35·Arena Elo + 0.30·benchmark average + 0.20·Intelligence Index + 0.15·LLM Stats, min-max normalized with missing-signal re-weighting. — see live data
Elo rating
A rating system borrowed from chess. In LLM arenas, models gain Elo by winning head-to-head human preference battles; the difference between two Elos predicts the win probability. — see live data
Benchmark
A standardized test set with scored tasks (math problems, coding issues, science questions) used to measure a specific model capability under comparable conditions. — see live data
Tokens
The sub-word chunks LLMs read and write. Roughly 1 token ≈ 4 characters of English text; pricing is quoted per million (M) tokens.
Context window
The maximum amount of text (in tokens) a model can consider at once — prompt plus answer. Frontier 2026 models range from 128K to 1M tokens. — see live data
TTFT (time to first token)
Latency until the first output token arrives. Dominates perceived responsiveness in chat; measured in milliseconds. — see live data
Output throughput (tokens/s)
How many tokens per second a model generates after the first token. Determines how fast long answers stream. — see live data
Mixture of Experts (MoE)
An architecture where only a subset of parameters (“experts”) activates per token, giving large knowledge at small inference cost — e.g. Qwen3.8 Max or DeepSeek V4. — see live data
Open weights
Models whose trained parameters are publicly released, allowing self-hosting and fine-tuning (Kimi K3, GLM-5.2, Llama 4). Contrast with proprietary API-only models. — see live data
Reasoning model
A model trained to spend extra “thinking” tokens before answering, trading latency for accuracy on hard problems (GPQA, HLE, competition math). — see live data
Hallucination
A confident but unsupported statement. Factuality benchmarks like SimpleQA measure how often a model avoids them. — see live data
SWE-Bench Verified
Resolving real GitHub issues in real repositories, verified by humans. The standard benchmark for agentic software engineering. — see live data
GPQA Diamond
Graduate-level “google-proof” science questions written by domain experts; a core reasoning benchmark. — see live data
Humanity’s Last Exam
An expert-authored question set designed to sit at the frontier of human knowledge; current frontier models still score in the 40–60% range. — see live data
MMLU-Pro
A hardened, 14-category successor to MMLU measuring broad academic knowledge with harder answer choices. — see live data
Min-max normalization
Rescaling any raw metric to 0–100 across the tracked model set, so scores from different sources can be blended. 80 always means “80% of the way from worst to best observed”. — see live data
Confidence interval (±CI)
The statistical uncertainty of an arena Elo. Overlapping CIs mean two models are effectively tied. — see live data
$/M tokens
Price per million tokens, quoted separately for input and output. Output tokens are typically 3–5× more expensive than input. — see live data