Articles for people who
ship the stack
Rewrites on React, Node.js, TypeScript and AI. A translated body is shown when it exists. If it does not, you see the English text and a note that says so.
Tagged: benchmarks
How Much Memory Do You Need for Local AI? Speeds from 16 GB to 512 GB
Measured local LLM speed and quality across 16, 32, 128, 256 and 512 GB machines, what each class can run, and which GPUs, Macs and mini PCs fit where.
5 min readMac Studio M5 Ultra vs DGX Spark: Local LLM Speed on the Same Models
Mac Studio M5 Ultra 256 GB against one, two and four DGX Spark boxes on Nemotron, Qwen3.8 and DeepSeek-V4.1-Flash: decode, prefill, batching and images.
4 min readRTX 5090 Local LLM Benchmarks: Same Qwen3.8-27B, 67 to 253 Tokens per Second
Qwen3.8-27B on one RTX 5090 across TensorFold, NInfer, ExLlamaV3, Qwentin and vLLM: decode, prefill, parallel throughput, quality and which stack to run.
4 min readRTX 5060 Ti 16 GB for Local AI: Measured LLM, Image and Speech Speeds
RTX 5060 Ti 16 GB tested for local AI: five LLMs from 42 to 116 tok/s, first-token times at 32K, image, video, 3D and speech speeds, and what to run on it.
6 min readTensorFold vs llama.cpp on a Mac Studio M5 Ultra: 1.84x the Throughput on Real JSON Work
I benchmarked TensorFold 1.2.0 against llama.cpp on a Mac Studio M5 Ultra with Qwen3.8-27B and 80 real JSON requests: 1.84x throughput at 8 concurrent.
9 min readHy-MT2-7B on Five Local Runtimes: Which One Should Serve Translation on a Mac?
Hy-MT2-7B load-tested on llama.cpp, LM Studio, mlx-lm, Ollama and MLX-Serve on two Mac Studio M5 Ultra machines: throughput, scaling, determinism, memory.
9 min readGemini 3.8 Flash at Launch: Speed, Pricing Deadline and Safety Caveats
A review of Gemini 3.8 Flash: official benchmark gains, a 304 tokens-per-second snapshot, promo pricing ending in 2027, the Cyber model and a multilingual safety dip.
7 min readWhy Dual RTX 3090s Hit 55 tok/s Instead of 220 with SGLang and DFlash2
A hands-on look at running Qwen3.8-27B on two 3090s with SGLang, and how PCIe x4 links, missing P2P, tight KV cache and power spikes crushed the advertised speed.
7 min readMiniCPM5-2B vs 4B Rivals: Why an Agentic Harness Tells a Different Story
MiniCPM5-2B beats 4B rivals on single-shot scores, then falls two to three times behind in agentic harnesses. The benchmark mix-up and Ollama throughput.
10 min read
About these articles
Request a 24h estimate
Need the same stack in a production operator layer? Send the brief — estimate within 24 hours.