ComputeArena
Base Compute Pty. Ltd. · Coding · Other
ComputeArena is a community benchmark platform for local AI. It collects AI throughput benchmarks from real machines, then ranks models by how fast they prefill prompts and decode tokens on specific chips. You install a CLI on macOS or Linux, run a model offline, and submit the signed report when you're ready to share it. The result is a benchmark leaderboard built from hardware people actually own, not a single lab's test bench.
About ComputeArena
What Is ComputeArena
ComputeArena answers a question that spec sheets can't: how fast does a given model actually run on the hardware you own? Instead of trusting vendor slides, it gathers benchmarks from volunteers who run models on their own laptops, desktops, and workstations. Every result links back to a signed report produced on real community hardware.
The platform sorts results by model, quantisation level, and chip, with separate numbers for prefill, which is how fast the model reads your prompt, and decode, which is how fast it writes the answer. That split matters, because a chip that handles long prompts well isn't always the one that streams tokens quickly.
Two things keep it from being a one-size answer. First, results depend on the exact runtime, quant, and context length used, so mismatched setups aren't directly comparable. Second, it focuses on throughput rather than answer quality, so it won't tell you which model writes better prose.
Getting Started
- Run the install command on macOS or Linux. The installer pulls the latest stable release and needs curl, tar, and Python 3. No sudo.
- Pick a runtime from the interactive menu. ComputeArena supports BaseRT and llama.cpp.
- Choose a model and a quantisation level, then start the benchmark. Everything runs offline.
- Wait for the run to finish, then review the saved report on your own machine.
- Sign in when you're ready to submit the benchmarks you want to make public.
Product Information
A quick look at ComputeArena's pricing, supported platforms, and performance.
Best for
The users, tasks, and scenarios where this tool fits best.
Users
- Local AI enthusiasts
- Hardware shoppers
- Model tinkerers
Tasks
- Measuring local LLM performance
- Choosing a quantisation level
- Validating a hardware upgrade
Scenarios
- Deciding whether a machine can host a model comfortably
- Comparing runtimes on one device
- Spotting outliers in a build
Key features
Prefill and decode split
ComputeArena reports two numbers for every run: PP512 prefill and TG128 decode. Prefill covers how fast the model reads your prompt, and decode covers how fast it writes back. Keeping them apart helps you see which half of the workload a chip handles better.
Community submitted benchmarks
Results come from volunteers rather than a single lab, so each entry links to a signed report generated on the contributor's own hardware, and the leaderboard shows who ran it and when. That gives you a spread of real setups instead of one controlled test bench.
Model, quant, and chip filters
The leaderboard lets you narrow results by model, runtime, quantisation, and chip. With hundreds of configurations listed across the whole site, the filters are the fastest way to find a comparison that matches your own machine, and they help you line up an entry against the chip you're thinking about buying.
Offline runs with opt-in sharing
The CLI runs the benchmark without an account, and reports stay on your disk until you decide to submit them. You can test as much as you like and only publish the runs you're happy with.
Runtime choice
You pick between BaseRT and llama.cpp when you start a run. Pop the same model through both. Running it twice shows how much the software stack, not just the silicon, changes your throughput.
Chip coverage on Apple and discrete GPUs
Entries cover Apple Metal chips and discrete GPUs through backends like Vulkan. That mix lets you compare an integrated Apple part against a dedicated graphics card on similar workloads.
Public contribution profiles
Contributors have profiles, so you can follow a specific person's submissions or check the history behind a result. It adds a bit of accountability to numbers that would otherwise float free.
Pros and cons
Pros
- Numbers come from real hardware across many contributors, so you get a realistic range instead of a single lab result.
- Splitting prefill and decode is more useful than one blended score.
- That's the pitch. No account needed to run benchmarks, and reports stay private until you submit them.
- Filtering by model, quant, and chip makes comparison straightforward.
Cons
- Results aren't strictly comparable unless runtime, quant, and context length match, so you have to check the setup behind each entry.
- It measures throughput only, which means there's no read at all on output quality, so it won't help you pick the smarter model for your task.
- Desktop only for now. macOS and Linux are supported, with no Windows path on the quickstart page.
Frequently asked questions
It's a community benchmark platform for local AI. People run models on their own machines and submit throughput results, which get ranked by model, quantisation, and chip. It's aimed at anyone who wants to know how fast a model runs on real hardware, whether that's a laptop on a desk or a workstation in a closet.
Related content
Explore related tools, skills, and articles for ComputeArena.
ComputeArena Alternatives
Base Code
Base44 (Wix) · Coding · ProductivityBase Code is a shared cloud development environment from Base44. Teammates describe changes to an existing GitHub codebase in chat, preview them live and send them back as pull requests.
Google EmbeddingGemma 2
Google DeepMind · Writing · Image · CodingGoogle EmbeddingGemma 2 is a 740M-parameter open model that turns text, code, images, audio, and video frames into vectors in one 768-dimension space, running on phones and laptops.
GitGlow
GitGlow · Coding · ProductivityGitGlow is a local-first desktop Git client that lets independent developers review their own code and their AI agents' changes in one place before shipping.