Institute for Holistic Technology Research GmbH
Holistech Labs
Hardware, inference engines and models, measured on our own machines and documented in full.
01 · Measurement series
Projects
Mac Studio M1 Ultra ·
Qwen3.8-Flash-Next with sushi on the Mac Studio M1 Ultra
Three builds of sushi on the same machine: releases 1.2.1 and 1.3.0 and a fork for the M1 Ultra, the same documents up to 245,000 tokens, two measurements per build.
Read the report →HP ZBook Ultra G1a ·
Qwen3.8-Flash-Next as a coding agent on the HP ZBook Ultra G1a
40 randomly drawn issues from SWE-bench Verified, two passes each, offline in OpenCode with halogen. What share gets solved, how long a pass takes and where the agent fails.
Read the report →Mac Studio M1 Ultra ·
Qwen3.8-Flash-Next as a coding agent on the Mac Studio M1 Ultra
40 randomly drawn issues from SWE-bench Verified, two passes each, offline in OpenCode with sushi on 64 GB. What share gets solved, how long a pass takes and where the agent fails.
Read the report →MacBook Pro M1 Max, Mac Studio M1 Ultra, HP ZBook Ultra G1a ·
Qwen3.8-Flash-Next as a coding agent in OpenCode
Six real issues from SWE-bench Verified, three passes each, offline: sushi on two 64 GB Macs and halogen on the HP ZBook Ultra G1a. How often the agent solves them, how long it takes and what else it tries.
Read the report →HP ZBook Ultra G1a ·
Long context with halogen 0.18
Qwen3.8-Flash-Next on halogen 0.18.0 from 32,000 to 990,000 tokens: how fast it reads and answers, and when it stops understanding.
Read the report →Mac Studio M1 Ultra ·
Qwen3.8-27B as an agent on the Mac Studio M1 Ultra
Six servers for Apple silicon with real agent histories: who keeps the history, how fast they read and write, and what speculation changes in the output and how much memory each engine takes at 128k.
Read the report →MacBook Pro M1 Max, Mac Studio M1 Ultra ·
Qwen3.8-Flash-Next on Apple M1 and Strix Halo
The same documents, questions and lengths up to 245,000 tokens: sushi on a MacBook Pro M1 Max and a Mac Studio M1 Ultra against halogen on the HP ZBook Ultra G1a.
Read the report →HP ZBook Ultra G1a ·
Long context with halogen 0.17 and checkpoint v2
Qwen3.8-Flash-Next on halogen 0.17.0 from 32,000 to 990,000 tokens: how fast it reads and answers, and when it stops understanding.
Read the report →HP ZBook Ultra G1a ·
Long context on the HP ZBook Ultra G1a
Qwen3.8-Flash-Next on halogen from 32,000 to 990,000 tokens: how fast it reads and answers, and when it stops understanding.
Read the report →HP ZBook Ultra G1a ·
Qwen3.8-Flash-Next on the HP ZBook Ultra G1a
Two engines, the same model, the same machine: llama.cpp against halogen-flash-server.
Read the report →HP ZBook Ultra G1a ·
Nine models on the HP ZBook Ultra G1a
One machine, nine local models: output speed, wait and reproducibility across three runs each.
Read the report →