local.ai is a benchmark site built to help you choose a configuration for running AI models locally. Its stated premise is that “running models locally is already easy; deciding what to run is the hard part.” Rather than scoring one variable at a time, it measures the full stack together: model, hardware, agent harness, inference engine, and quantization method. The speed and evaluation data are produced by Exo Labs, itself a player in local inference, while the site describes itself as “an independent reference with no commercial relationships to hardware vendors or model providers.” We will come back to this arrangement later in the article.
Key Features
- Measurement at the system-combination level: Instead of benchmarking a model alone or a piece of hardware alone, it treats a single configuration ─ model, hardware, inference engine, and even quantization or pruning ─ as one unit of measurement
- Three metrics: intelligence, speed, and cost: “Intelligence” is an Intelligence Index derived from a weighted average across evaluation suites such as GDPVal; “speed” is end-to-end task completion time; “cost” is anchored on hardware price and electricity
- A deliberate decision not to make token throughput the headline metric: In one published example, two configurations with nearly identical decode speeds (17.77 tok/s and 17.14 tok/s) show median GDPVal task completion times of 39.1 minutes and 11.9 minutes. The site uses real measurements to argue that picking a configuration on tok/s alone does not hold up
- Hardware frontier analysis: Real machines such as the Mac M4 Max (128GB, 40-core GPU) and NVIDIA’s DGX Spark (GB10, 128GB) are plotted by price tier on a combined intelligence-and-speed axis. You can read off findings like this: on memory bandwidth the M4 Max leads at 546 GB/s versus 273 GB/s for the DGX Spark, a 2x gap, yet median task completion time narrows to 19.8 minutes versus 28.7 minutes
Pricing
Browsing the benchmarks and creating a profile are free. The public pages do not mention any paid plan, so if you need one, checking the official site is the reliable route.
Pros and Cons
✅ Pros
- Differences you cannot read from a spec sheet ─ such as “same tok/s, but 3x the completion time” ─ are presented as numbers, so you can narrow down candidates before buying
- Because listings include quantization as part of the combination, you can judge whether there is room to improve simply by changing the configuration on the machine you already own, without replacing it
- Since entries are sorted by configuration rather than model name, you can make the comparison people actually agonize over: is a large model with heavy quantization or a small model with light quantization the better deal?
⚠️ Cons
- While it claims neutrality, the measurements are taken by Exo Labs, a company whose business is local inference, and there is no published mechanism for third parties to reproduce the same numbers. This tension is one that any benchmark calling itself independent always carries
- It launched only in August 2026, so how thoroughly configurations will be covered and how often the data will be refreshed cannot yet be judged from a track record
Comparison with Similar Services
| Criteria | local.ai | Artificial Analysis | MLPerf |
|---|---|---|---|
| Primary focus | Local execution (model + hardware configuration) | Models served via cloud APIs | Inference and training performance submitted by vendors |
| Speed metric | Real-task completion time | Primarily token throughput | Standardized benchmarks |
| Cost axis | Hardware price + electricity | API pricing | Performance is the main concern |
| Who measures | The operator, first-hand | The operator, via API | Vendors submit, results are audited |
The division of labor is clear: Artificial Analysis for cross-comparing cloud APIs, MLPerf for official vendor figures, and local.ai when you plan to run things on your own machine.
Who Should Use It
- People considering a GPU or chip-equipped machine for local LLMs who want decision criteria beyond tok/s
- People keeping their current machine who want to find out whether a different model-and-quantization combination could improve results
- People who want to compare on cost whether cloud API spending could be replaced by a hardware purchase
Summary
This is a site for finding a “fast configuration” rather than a “fast model.” At its core is a design that presents, in measured values, the completion-time differences that tok/s cannot reveal, which makes it well suited to narrowing candidates before a purchase. As noted above, the measurements come from an interested party, so the practical way to read it is not as rankings to accept at face value but as an explanation of why a given configuration is fast.