STORY RECORD
ExoLabs founder Alex Cheema opens early access to local.ai—a pareto-frontier benchmark map for 1,000+ local-AI hardware setups, born at Apple Park and built on NVIDIA-donated hardware
Alex Cheema, the ExoLabs founder and former Oxford researcher, is rolling out early access to local.ai, an independent benchmarking website that maps which local-AI hardware (MacBook, Mac Mini, DGX Spark, clusters) runs which models at which quantization and multi-token-prediction setting.
Cheema says the idea originated at Apple Park and was built at NVIDIA HQ using hardware NVIDIA donated without strings attached.
The site benchmarks over 1,000 unique configurations with what Cheema describes as end-to-end real agent harnesses and displays results in pareto-frontier rankings.
Access is being distributed in batches; Cheema posted July 16 that he was sending out the second batch of codes in a few hours. The site was initially previewed around July 2 at the AI Engineer World's Fair, where Cheema and collaborator Sero discussed it on the ThursdAI podcast.
The claims about methodology, dataset size, and NVIDIA's no-strings donation are currently single-source from Cheema's post.
Why It Matters
Local AI is at an inflection point—Apple is signaling 1.5 TB unified-memory chips with the M7 Ultra, NVIDIA is pushing the DGX Spark, and the open-weight model ban debate is intensifying.
The community has no single reliable source to answer the most common hardware-buying question: what hardware gives the best real-world inference for a given local model? Cheema's local.ai directly targets that gap with a free, independent, pareto-frontier approach.
If the methodology holds up, it becomes the default buyer's guide for the local-AI era. If it is a marketing funnel for ExoLabs, the community needs to know that too.
The timing—right after Cheema's warning that 1TB VRAM is the floor for frontier local AI—makes this a practical sequel: here is the tool to figure out what setup actually fits your budget.
The Facts
8Alex Cheema posted on July 16, 2026, announcing early access to local.ai, described as an independent benchmarking website for local AI.
strong · confidence 0.95
Cheema states the idea for local.ai originated at Apple Park and was built at NVIDIA HQ using hardware NVIDIA donated with 'no strings attached.'
weak · confidence 0.45
Cheema claims local.ai benchmarks over 1,000 unique setups using end-to-end real agent harnesses, covering every hardware, model, quantization, and MTP setting, and maps results into pareto frontiers.
weak · confidence 0.35
The site directly answers three questions Cheema says Apple and NVIDIA field from customers: what models can run locally on a Mac/Spark, which Mac or DGX Spark to buy, and whether machines can be clustered for bigger models.
weak · confidence 0.6
Access to local.ai is being rolled out in batches. Cheema posted that he was sending out codes for the second batch a few hours after his July 16 post.
strong · confidence 0.9
Cheema is the founder of ExoLabs, describes himself as 'prev @UniOfOxford', has approximately 51,850 followers, and is verified with a blue check on X.
strong · confidence 0.95
As of approximately 30 minutes after posting, Cheema's announcement had 54 likes, 5 retweets, 17 replies, 0 quotes, 2 bookmarks, and 1,404 views. Approximately 90 minutes later, metrics grew to 63 likes, 6 retweets, 18 replies, 1 quote, 4 bookmarks, and 1,743 views.
strong · confidence 0.95
The post includes three attached photos (dimensions 1536x2048 each), suggesting screenshots or images of the local.ai interface.
strong · confidence 0.9
Still Open
5- OpenThe local.ai website itself was not fetched or independently verified as live and functional. The claim of an existing site with 1,000+ benchmarked setups rests entirely on Cheema's post.
- OpenNVIDIA's 'no strings attached' hardware donation is a single-source claim from Cheema. No NVIDIA spokesperson or independent source has confirmed the arrangement or the terms.
- OpenThe benchmarking methodology ('end-to-end with real agent harnesses') is described but not independently audited. The specific models, hardware configurations, and quantizations included are not listed.
- OpenCheema's claim that Apple Park was the birthplace of the idea is a narrative detail that cannot be independently corroborated. Cheema attended Apple's Local AI event at Apple Park on June 23, 2026, making the setting plausible but not confirmed as the origin point.
- OpenThe total number of unique setups (1,000+) and the claim of covering 'every hardware, every model, every quantization, every MTP setting' are non-specific boasts typical of launch announcements. The actual coverage scope is unclear.