openbench

Provider-agnostic, open-source evaluation infrastructure for language models

openbench provides standardized, reproducible benchmarking for LLMs across 30+ evaluation suites (and growing) spanning knowledge, math, reasoning, coding, science, reading comprehension, health, long-context recall, graph reasoning, and first-class support for your own local evals to preserve privacy. Works with any model provider - Groq, OpenAI, Anthropic, Cohere, Google, AWS Bedrock, Azure, local models via Ollama, Hugging Face, and 30+ other providers.

To get started, see the tutorial below or reference the docs.

Features

• 🎯 95+ Benchmarks: MMLU, GPQA, HumanEval, SimpleQA, competition math (AIME, HMMT), SciCode, GraphWalks, and more
• 🔧 Simple CLI: bench list, bench describe, bench eval (also available as openbench), -M/-T flags for model/task args, --debug mode for eval-retry, experimental benchmarks with --alpha flag
• 🏗️ Built on inspect-ai: Industry-standard evaluation framework
• 📊 Extensible: Easy to add new benchmarks and metrics
• 🤖 Provider-agnostic: Works with 30+ model providers out of the box
• 🛠️ Local Eval Support: Privatized benchmarks can be run with bench eval <path>
• 📤 Hugging Face Integration: Push evaluation results directly to Hugging Face datasets

🏃 Speedrun: Evaluate a Model in 60 Seconds

Prerequisite: Install uv

# Create a virtual environment and install openbench (30 seconds)
uv venv
source .venv/bin/activate
uv pip install openbench

# Set your API key (any provider!)
export GROQ_API_KEY=your_key  # or OPENAI_API_KEY, ANTHROPIC_API_KEY, etc.

# Run your first eval (3 seconds)
bench eval mmlu --model groq/openai/gpt-oss-120b --limit 10

# That's it! 🎉 Check results in ./logs/ or view them in an interactive UI:
bench view

openbench.mp4

Supported Providers

openbench supports 30+ model providers through Inspect AI. Set the appropriate API key environment variable and you're ready to go:

Provider	Environment Variable	Example Model String
AI21 Labs	AI21_API_KEY	ai21/model-name
Anthropic	ANTHROPIC_API_KEY	anthropic/model-name
AWS Bedrock	AWS credentials	bedrock/model-name
Azure	AZURE_OPENAI_API_KEY	azure/<deployment-name>
Baseten	BASETEN_API_KEY	baseten/model-name
Cerebras	CEREBRAS_API_KEY	cerebras/model-name
Cohere	COHERE_API_KEY	cohere/model-name
Crusoe	CRUSOE_API_KEY	crusoe/model-name
DeepInfra	DEEPINFRA_API_KEY	deepinfra/model-name
Friendli	FRIENDLI_TOKEN	friendli/model-name
Google	GOOGLE_API_KEY	google/model-name
Groq	GROQ_API_KEY	groq/model-name
Helicone	HELICONE_API_KEY	helicone/model-name
Hugging Face	HF_TOKEN	huggingface/model-name
Hyperbolic	HYPERBOLIC_API_KEY	hyperbolic/model-name
Lambda	LAMBDA_API_KEY	lambda/model-name
MiniMax	MINIMAX_API_KEY	minimax/model-name
Mistral	MISTRAL_API_KEY	mistral/model-name
Moonshot	MOONSHOT_API_KEY	moonshot/model-name
Nebius	NEBIUS_API_KEY	nebius/model-name
Nous Research	NOUS_API_KEY	nous/model-name
Novita AI	NOVITA_API_KEY	novita/model-name
Ollama	None (local)	ollama/model-name
OpenAI	OPENAI_API_KEY	openai/model-name
OpenRouter	OPENROUTER_API_KEY	openrouter/model-name
Parasail	PARASAIL_API_KEY	parasail/model-name
Perplexity	PERPLEXITY_API_KEY	perplexity/model-name
Reka	REKA_API_KEY	reka/model-name
SambaNova	SAMBANOVA_API_KEY	sambanova/model-name
SiliconFlow	SILICONFLOW_API_KEY	siliconflow/model-name
Together AI	TOGETHER_API_KEY	together/model-name
Vercel AI Gateway	AI_GATEWAY_API_KEY	vercel/creator-name/model-name
W&B Inference	WANDB_API_KEY	wandb/model-name
vLLM	None (local)	vllm/model-name

Available Benchmarks

See the Benchmarks Catalog or use bench list.

Commands and Options

For a complete list of all commands and options, run: bench --help See the docs for more details.

Command	Description
bench list	List available benchmarks
bench eval <benchmark>	Run benchmark evaluation
bench eval-retry <log_files>	Retry a failed evaluation
bench view	Interactive UI to view benchmark logs
bench cache <info/ls/clear/upload>	Manage OpenBench caches

Common eval Configuration Options

Option	Environment Variable	Default	Description
-M <args>	None	None	Pass provider/model-specific arguments (e.g., -M only=groq)
-T <args>	None	None	Pass task-specific arguments to the benchmark
--model	BENCH_MODEL	groq/openai/gpt-oss-20b	Model(s) to evaluate
--epochs	BENCH_EPOCHS	1	Number of epochs to run each evaluation
--epochs-reducer	BENCH_EPOCHS_REDUCER	None	Reducer(s) applied when aggregating epoch scores
--max-connections	BENCH_MAX_CONNECTIONS	10	Maximum parallel requests to model
--temperature	BENCH_TEMPERATURE	0.6	Model temperature
--top-p	BENCH_TOP_P	1.0	Model top-p
--max-tokens	BENCH_MAX_TOKENS	None	Maximum tokens for model response
--seed	BENCH_SEED	None	Seed for deterministic generation
--limit	BENCH_LIMIT	None	Limit evaluated samples (number or start,end)
--logfile	BENCH_OUTPUT	None	Output file for results
--sandbox	BENCH_SANDBOX	None	Environment to run evaluation (local/docker)
--timeout	BENCH_TIMEOUT	10000	Timeout for each API request (seconds)
--fail-on-error	None	1	Threshold of allowable sample errors (use an integer for count or a float for proportion)
--display	BENCH_DISPLAY	None	Display type (full/conversation/rich/plain/none)
--reasoning-effort	BENCH_REASONING_EFFORT	None	Reasoning effort level (low/medium/high)
--json	None	False	Output results in JSON format
--log-format	BENCH_LOG_FORMAT	eval	Output logging format (eval/json)
--hub-repo	BENCH_HUB_REPO	None	Push results to a Hugging Face Hub dataset
--keep-livemcp-root	BENCH_KEEP_LIVEMCP_ROOT	False	Allow preservation of root data after livemcpbench eval runs
--code-agent	BENCH_CODE_AGENT	opencode	Select code agent for exercism tasks

Development and Building Your Own Evals

For a full guide, see Contributing Guidelines and Extending openbench. Also, check out Inspect AI's excellent documentation.

Quick Eval: Run from Path

For one-off or private evaluations, point openbench directly at your eval:

bench eval /path/to/my_eval.py --model groq/llama-3.3-70b-versatile

Plugin System: Distribute as Packages

openbench supports a plugin system via Python entry points. Package your benchmarks and distribute them independently:

# pyproject.toml
[project.entry-points."openbench.benchmarks"]
my_benchmark = "my_pkg.metadata:get_benchmark_metadata"

After pip install my-benchmark-package, your benchmark appears in bench list and works with all CLI commands. Perfect for:

• Sharing benchmarks across teams
• Versioning evaluations independently
• Overriding built-in benchmarks with custom implementations

FAQ

How does openbench differ from Inspect AI?

openbench provides:

• Reference implementations of 20+ major benchmarks with consistent interfaces
• Shared utilities for common patterns (math scoring, multi-language support, etc.)
• Curated scorers that work across different eval types
• CLI tooling optimized for running standardized benchmarks

Think of it as a benchmark library built on Inspect's excellent foundation.

Why not just use Inspect AI, lm-evaluation-harness, or lighteval?

Different tools for different needs! openbench focuses on:

• Shared components: Common scorers, solvers, and datasets across benchmarks reduce code duplication
• Clean implementations: Each eval is written for readability and reliability
• Developer experience: Simple CLI, consistent patterns, easy to extend

We built openbench because we needed evaluation code that was easy to understand, modify, and trust. It's a curated set of benchmarks built on Inspect AI's excellent foundation.

How can I run bench outside of the uv environment?

If you want bench to be available outside of uv, you can run the following command:

uv run pip install -e .

I'm running into an issue when downloading a dataset from HuggingFace - how do I fix it?

Some evaluations may require logging into HuggingFace to download the dataset. If bench prompts you to do so, or throws "gated" errors, defining the environment variable

HF_TOKEN="<HUGGINGFACE_TOKEN>"

should fix the issue. The full HuggingFace documentation can be found on the HuggingFace docs on Authentication.

See the docs for further Tips and Troubleshooting.

🚧 Alpha Release

We're building in public! This is an alpha release - expect rapid iteration. The first stable release is coming soon.

Quick links:

• Report a bug
• Request a feature

Reproducibility Statement

As the authors of openbench, we strive to implement this tool's evaluations as faithfully as possible with respect to the original benchmarks themselves.

However, it is expected that developers may observe numerical discrepancies between openbench's scores and the reported scores from other sources.

These numerical differences can be attributed to many reasons, including (but not limited to) minor variations in the model prompts, different model quantization or inference approaches, and repurposing benchmarks to be compatible with the packages used to develop openbench.

As a result, openbench results are meant to be compared with openbench results, not as a universal one-to-one comparison with every external result. For meaningful comparisons, ensure you are using the same version of openbench.

We encourage developers to identify areas of improvement and we welcome open source contributions to openbench.

Acknowledgments

This project would not be possible without:

• Inspect AI - The incredible evaluation framework that powers openbench
• EleutherAI's lm-evaluation-harness - Pioneering work in standardized LLM evaluation
• Hugging Face's lighteval - Excellent evaluation infrastructure

Citation

@software{openbench,
  title = {openbench: Provider-agnostic, open-source evaluation infrastructure for language models},
  author = {Sah, Aarush},
  year = {2025},
  url = {https://openbench.dev}
}

License

MIT

---

Built with ❤️ by Aarush Sah and the Groq team