Alex Cheema
13 Intel
Alex Cheema is a London-based engineer and co-founder of EXO Labs, a startup founded in 2024 focused on distributed computing for running AI models on everyday consumer hardware. He studied physics at the University of Oxford and previously worked on DeFi trading infrastructure and video live-streaming systems, including as CTO of GamingForGood (now TSG).
Is this you? Sign in with X to claim this profile.
Intel
Alex Cheema sketches running Kimi K3 locally: four 512GB M3 Ultra Mac Studios, with an expected 30-plus tokens per second using MTP and tensor parallelism over Thunderbolt 5 RDMA.
Alex Cheema previews Qwen 3.8 27B's weight release: close to GLM-5.2-level intelligence running locally on a $4k 36GB MacBook at roughly 25 tokens per second.
Apple is now marketing four M5 Ultra Mac Studios clustered over Thunderbolt 5 RDMA as a local frontier-model rig — 2TB of unified memory at 4.8TB/s, pitched to run trillion-parameter models on-prem — with exo's software as the clustering layer and the setup featured on Apple's own product pages.
Exo Labs' founder marks the M5 Ultra Mac Studio as the moment clustered consumer Macs reach frontier-class local inference — roughly 1,000x the two-MacBook Llama 405B run, compounding a 10x hardware step with RDMA over Thunderbolt, better kernels and denser models.
A case for splitting inference across mismatched local hardware: prefill is compute bound and runs faster on a DGX Spark, decode is memory bandwidth bound and runs faster on an M5 Max, so run each phase where it is fastest and ship the KV cache between them.
Alex Cheema runs the local-inference maths on a freshly published 2.4T-parameter, 95B-active open model: 4.89TB of weights, no 4-bit quant from the lab, and four 512GB M3 Ultras as the only viable local 4-bit rig — with an NVFP4 quant from NVIDIA, Unsloth or Red Hat the missing piece.
Argues that agents able to benchmark their own kernels are quietly ending CUDA lock in.
The full vLLM configuration for running Nemotron 3.5 Lightning on a DGX Spark with DSpark, covering the mixture-of-experts backend, FP8 KV cache, 262K context length and Mamba cache settings.
A working note on multi-token prediction and speculative decoding as near-free throughput for local models: acceptance rates climb with structured tokens so coding gains most, MLX has been slow to add MTP, and MTPLX already ships it on Mac ahead of Qwen 3.8 27B's weights.
Alex Cheema places Nemotron 3.5 Lightning firmly on the intelligence-versus-speed Pareto frontier, displacing Holo 3.1 35B A3B, the post-trained Qwen variant that previously held that region.
Alex Cheema compares Nemotron 3.5 Lightning against Qwen 3.6 35B A3B on DGX Spark: Qwen decodes faster per token but is 45% more verbose, so tasks take 42% longer overall.
Alex Cheema measures Nemotron 3.5 Lightning completing tasks 5.36x faster than the most intelligent DGX Spark-capable model while retaining 78% of its intelligence, with the speedup coming almost entirely from per-token decode speed.
Cheema questions why more than 10 specialized LLM inference engines have shipped in a month when vLLM and sgLang already exist, drawing replies debating whether this is healthy competition or harmful ecosystem fragmentation.