Builder rig interviews
Five short questions, answered in writing by people who run the models on hardware they own. Speeds are theirs, labelled self-reported.
Dual Intel Arc Pro B70 (64GB) local LLM: vLLM and llama.cpp (Josh Gold)
Josh Gold runs a dual Intel Arc Pro B70 (64 GB) local LLM rig: vLLM on one card, llama.cpp Vulkan on the other.
@c4software's two 128 GB Strix Halo boxes
@c4software runs Qwen3.8-27B on two 128 GB Ryzen AI Max+ 395 boxes: llama.cpp on bigchuck, gufo on silentchuck.
RTX 5090 32 GB + RTX 3090 24 GB Workstation Running q27 (Quasar) & Whisper Large v3
Gabe Ortiz runs q27 (Quasar) on an RTX 5090 and Whisper large v3 on an RTX 3090 in one Ryzen 9 9900X workstation.
Fabio Matricardi's ThinkPad 780M CPU AI lab
Fabio Matricardi runs llama.cpp Vulkan on a ThinkPad AMD 780M and a Dell Intel Ultra iGPU, both with 32 GB unified RAM.
James Meadlock's GhettoRak: eight DGX Spark nodes
James Meadlock's GhettoRak is eight DGX Spark nodes (1,024 GB) running GLM-5.3 and DeepSeek V4.1 Flash.
Xijing (Thomas) Wang's MacBook Pro M5 Max 128 GB
Xijing (Thomas) Wang serves Qwen3.8-27B at 5-bit with oMLX on a MacBook Pro M5 Max with 128 GB unified memory.
Roshan George's RTX A6000 48 GB shared vLLM server
Roshan George shares one used RTX A6000 48 GB, capped at 250 W, between vLLM and camera decode on an AMD EPYC server.
NVIDIA DGX Spark (128GB) local LLM: Beam-501B and Qwen3.8-Flash-Next tokens per second with ds4-dfm-rs (Sunghoon Baek, @Baekpica)
Sunghoon Baek runs Beam-1.0-501B and Qwen3.8-Flash-Next on one 128 GB DGX Spark with ds4-dfm-rs. Local LLM, self-reported.
Danila Sukharev (@ddanila)'s 36 GB MacBook Pro M3 Pro: Local Qwen OCR, Weird GPT-2 Text & Building llm-lang-lab
Danila Sukharev (@ddanila) runs Qwen-class models for OCR and GPT-2 on a 36 GB M3 Pro MacBook, and builds llm-lang-lab. Self-reported.
Danylo Chernov (@DanielChernov7, NoctHost): What Happens When You Run 3B–14B Ollama Models on Ordinary Cloud CPUs
Danylo Chernov (NoctHost) timed 5 Ollama Q4_K_M models on Hetzner CX23, CX33 and CCX23. A 720-token prompt took 40 seconds to first word. Self-reported.
OscarJHernandez's GMKtec EVO-X2 128 GB: models, quants, and a benchmark still being analyzed
OscarJHernandez experiments with models and quants on a 128 GB GMKtec EVO-X2. The multi-day benchmark is still being analyzed. Self-reported.
Framework Desktop (128GB) Ryzen AI Max+ 395 Strix Halo LLM benchmark (Valérian de Thézan de Gaussan)
Framework Desktop 128GB Ryzen AI Max+ 395 Strix Halo LLM benchmark by Valérian de Thézan de Gaussan. Self-reported, 28 Oct 2026.
Questions
- What is a builder rig interview?
- Five short questions, answered in writing by people who run open models on hardware they own. Speeds are theirs, labelled self-reported. They are community measurements, not ours.
- Which builder rig interviews are published?
- 12 interviews: Josh Gold, @c4software, Gabe Ortiz, Fabio Matricardi, James Meadlock, Xijing (Thomas) Wang, Roshan George, Sunghoon Baek, Danila Sukharev, Danylo Chernov, OscarJHernandez, Valérian de Thézan de Gaussan. Each one links to the models and the hardware named in the answers.
- Can I test open models in the browser before buying hardware?
- Yes. The free browser test is at https://localllmrig.com/.