We introduce MARL‑BattleGrounds (MARL‑BGs), a JAX‑native, user‑friendly benchmark tailored for heterogeneous and competitive multi‑agent reinforcement learning (MARL). It balances sample efficiency, computational efficiency, and researcher‑centric design. Existing MARL benchmarks tend to be either very simple and fast like SMAX, where victory reduces to unit micromanagement, or extremely rich and costly like Dota 2. MARL‑BGs fills this missing middle ground.\ \ The first mode, Team Deathmatch, is a simultaneous turn‑based game between two teams of five heterogeneous agents. The five classes each have distinct basic, passive, and ultimate abilities that synergize with allies and counter opponents. We provide 52 hand‑crafted maps, including a 12‑map curriculum and ten held‑out validation and test maps.\ \ MARL‑BGs offers two canonical evaluation schemes. The first consists of eight end‑game scenarios that form a hard generalization challenge: each scenario asks a team to turn a nearly lost position into a win against a strong reactive controller within 5‑10 ticks. The second is a monthly round‑robin tournament that retrains each submission on equal compute, pits it against every member of the current “Big Nine” field, and publishes Elo results on the official leaderboard; all entrants are released as downloadable opponents.\ \ Each game logs 11 192 metrics across 28 topics. The benchmark also includes three strong scripted opponents, an analytical replay viewer, and off‑the‑shelf tools for self‑play, cross‑play, curriculum learning, and zero‑shot coordination evaluation. An API for LLM‑based agents is provided, and we use it to evaluate Qwen3.5‑9B against our strongest policy. Through the BattleClient, humans can play alongside or against any mix of policies, enabling human‑AI evaluation. We benchmark MARL‑BGs with six additional deep MARL baselines, all of which are released. On an RTX 5090 running 1 024 parallel games, the simulator achieves 50 000 five‑vs‑five game steps per second.\ \ Review: MARL‑BGs delivers a high‑performance yet richly featured environment that bridges the gap between minimalist and heavyweight MARL benchmarks, making it an ideal testbed for research on heterogeneous teamwork, zero‑shot coordination, and the integration of LLMs with reinforcement learning.