Alternatives to BALROG

Open-source toolkit for benchmarking vision-language models across 80+ multimodal tasks, with rankings published on the Open VLM Leaderboard.

Open-source benchmark that scores AI agents on real, multi-step software-company work tasks.

Open-source benchmark that runs AI browser agents on live websites and scores the HTTP request they actually fire.

OpenCompass (司南) benchmarks LLMs, VLMs, and AI4S models on 100+ open evaluation datasets with dated leaderboards

ViDoRe Benchmark: an open visual document retrieval benchmark and ColPali model suite for enterprise RAG evaluation.

AppWorld is a simulated-world benchmark for evaluating AI coding agents across 9 apps and 457 APIs.

VisualWebArena is a 910-task benchmark from Carnegie Mellon that scores multimodal browser agents on whether they actually complete visual web tasks.

Chronos Engine (Reality Engine) is an open-source AI temporal sandbox for simulating alternate histories and cause-effect cascades across

Hume AI measures 600+ dimensions of expression in voice across 50+ languages, so you can test whether a voice agent was understood, not just whether it

An independent AI leaderboard ranking 408 models on one LLM Stats Score that blends benchmarks with live API speed and price.

Free public leaderboard benchmarking LLMs on competition math and Lean proof benchmarks like ArXivMath and ArXivLean.

Free Hugging Face agent leaderboard ranking LLMs on planning, tool use, and multi-step agentic tasks

PandaProbe is open-source agent observability that traces runs, evaluates outcomes, and turns failures into validated fixes via its Repair

GameEngineBench scores frontier AI models on real game-engine programming — see which model can actually ship UE5 code.

Free, browser-based AI benchmark where models build 3D voxel structures from text prompts and humans vote on the results.

WYRD ECS Core is an open-source blueprint for a real-time entity-graph world model on an Entity-Component-System architecture.

Open-source interactive world model that turns keyboard and mouse input into real-time 25 fps video without a game engine.

Silico is Goodfire's interpretability agent for understanding and debugging the internals of your AI models.

Arena AI ranks chat models, coding agents and fullstack builds on live head-to-head human votes.

Fiddler AI is an enterprise AI control plane for agent observability, inline guardrails, and auditable AI governance.

Weights & Biases tracks ML experiments and traces LLM apps so teams ship AI models faster

Patronus AI builds simulated environments and training data for reinforcement learning on frontier agent models.

Paste one prompt, see how 40+ AI models answer it, then pick the one that's best, fastest, or cheapest.

Odyssey-2 Max is a causal world model that simulates physical scenes in real time as you feed it actions.

Polymath builds simulation environments for training and evaluating autonomous AI agents on long-horizon, multi-tool work

Simulate emergent societies with autonomous AI agents in a multi-agent world engine.

MarioGPT turns plain-English prompts into playable Super Mario Bros levels using a fine-tuned GPT-2 text2level model.