Twitxy
Saved
@halfpennymac's profile picture
View on X

HalfpennyMac

Verified account This account is verified. Learn more

@halfpennymac

Dedicated Mac minis, rented monthly. UK-based Apple silicon for iOS builds, AI workloads and browser automation.
Cotswolds, UK halfpennymac.com Joined May 2026
170 Following 113 Followers 100 Posts

Latest posts 5

HalfpennyMac Verified account @halfpennymac
Nanbeige4.3-33B's model card puts it ahead of Qwen3.8-27B at agentic coding. We ran both on one 48GB M5 Pro Mac mini with our own tests.

On short coding and maths tasks they came out close, Qwen slightly ahead and faster at long context.

Numbers: halfpennymac.com/nanbeige43-33b…
halfpennymac.com Nanbeige4.3-33B on a 48GB Mac mini: speed, quality and the context bill Real llama.cpp numbers for Nanbeige4.3-33B on a 48GB Mac mini: speed per quant, KL divergence, HumanEval and GSM8K vs Qwen3.8-27B, and the KV-cache bill.
HalfpennyMac Verified account @halfpennymac
Every Halfpenny Mac rack has a little HP box watching over it, and it takes the job very seriously. It tracks UPS power, battery levels and network latency around the clock, and the moment anything wobbles, the alerts go out. It's never had a day off, and it's never once asked for one.
HalfpennyMac Verified account @halfpennymac
PrismML dropped Bonsai 2 27B this month: a 27B model in a 5.95GB file. We squeezed it onto a Mac mini with 8GB of unified memory to see if it runs. It does: 7.6 tok/s, every layer on the GPU, real working code. The four settings that make it fit (and the one that crashed the machine): halfpennymac.com/bonsai-2-27b-8…
halfpennymac.com Bonsai 2 27B on an 8GB Mac mini: it runs, and here is where it stops A 27B model on an 8GB Mac: real llama.cpp numbers from an M2 Mac mini, the settings that make it fit, and the one that crashed it.