HMIC / Independent research

We characterize model chains.

We test the whole chain on our own machines and publish what we find, with the receipts.

01

Where to go

02

Most watched

Qwen3.6-35B Q4 versus Q6 benchmark video thumbnail

The Assembled Worker / Video / September 2026

Qwen3.6-35B: Q4 versus Q6

In our installed comparison, Q4 ran nearly three times faster while retaining about 98% of the measured quality. Why the useful quant is not always the largest one.

Watch
Qwen3.6-35B versus Qwen3.8-27B decode speed video thumbnail

The Assembled Worker / Video / September 2026

Four times the decode speed, and we still start with the 27B

Qwen3.6-35B decodes about four times faster than the dense Qwen3.8-27B. Why speed alone did not pick our starting model.

Watch
Qwen3-0.6B accuracy fix video thumbnail

The Assembled Worker / Video / September 2026

Qwen3-0.6B: one fix, half the questions back

The same 200 questions, before and after one fix. A tiny model goes from almost nothing right to about half.

Watch
Cover of Canton Sea by TJ Jung

Book / September 2026

Canton Sea

TJ Jung's first nonfiction book develops a practical theory of human-AI signal-chain identity across thirty-eight short chapters. Paperback and Kindle.

Open the book page
LTX-2.5 prompting tips video thumbnail

Video / August 2026

LTX-2.5: prompting tips from 100 controlled tests

A hundred controlled video generations, and the prompting fixes that actually changed the output.

Watch
Pixal3D versus TRELLIS.2 image-to-3D video thumbnail

Video / September 2026

Pixal3D vs TRELLIS.2: beyond the pretty preview

120 controlled image-to-3D tests, judged on the model you actually get, not the first render.

Watch
03

Open source tools

Bounded JSON Input

A strict Python boundary for JSON input. It rejects duplicate keys, invalid UTF-8, nonfinite numbers and oversized or deeply nested documents.

Read about the release

Process Haltere

A status calculator for long-running work. It uses durable progress timestamps to flag stalls and makes its limits explicit.

Read about the release
04

How we work

We run it ourselves

A named file, its runtime and settings, on our own hardware.

We write the missing datasheet

Each run is recorded like a certificate of analysis for that exact file.

We state the boundary

Every study says what it does not prove. Corrections are dated, not rewritten.

05

Founder

TJ Jung founded HMIC and does the work on these pages. He is a systems architect and Senior AI Implementation Engineer specializing in pipeline automation and model routing, and the author of Canton Sea.

Surfer, lead guitar player, ex-pizza guy, eleven years in biologics, engineer. The habit is the same in each: trace the signal through an imperfect chain, isolate the variable that controls the outcome, and keep what works.

Contact