Local AI ranking
Part of one ranking: which AI runs best on the machine you already own

10 app tasks.
Every machine class.
Nobody touched the code.

Every contender got the exact same 10 prompts and worked alone in a coding agent: wrote the code, ran it in a real browser, fixed its own bugs. Then an automatic checker clicked through every app. Round 1 gave 45 minutes per app; Round 2 gives up to 150 and tests recipes by @MiaAI_lab and @ashxhart.

133apps you can click
2rounds
4machine classes
0cloud
Round 1 · 29 Sep 2026

10 apps, 5 setups, 45 minutes each

The opening night: five local setups on four machine classes. 48 apps, all results below.
Jump to Round 1 ↓Task × setup matrix
Round 2 · 30 Sep – 03 Oct 2026

9 cells, recipes by Mia and Ash

Same 10 tasks, up to 150 minutes for the three big ones, tiers 0–3 per task. Each class unlocks with its result post — nothing is on this page before its post.
16 GB live32 GB live128 GB live256 GB liveNotes & checks
16 GB class · contender 1

Qwen3.8 27B on RTX 5060 Ti, 16 GB

llama.cpp · Unsloth UD-Q3_K_XL · 96K context · KV cache q4_0 · MTP 2

10 / 10apps passed every automatic check
14 minmedian until the first working version
45 minper app, alone, no human help
Model fit calculator built by Qwen3.8 27BOpen app ↗
1 · Model fit calculator
Will this AI model fit on my graphics card?
All 16 checks passedFirst working version after 8 min
3D scene built by Qwen3.8 27BOpen app ↗
2 · 3D scene
The home AI lab at night, in real-time 3D
All 13 checks passedFirst working version after 12 min
Product page built by Qwen3.8 27BOpen app ↗
3 · Product page
A landing page for an invented home AI box
All 17 checks passedFirst working version after 10 min
Platformer game built by Qwen3.8 27BOpen app ↗
4 · Platformer game
A 2D jump and run with 3 levels
All 18 checks passedFirst working version after 15 min
Time tracker built by Qwen3.8 27BOpen app ↗
5 · Time tracker
Time tracking with project billing
All 15 checks passedFirst working version after 19 min
Dashboard built by Qwen3.8 27BOpen app ↗
6 · Dashboard
Local AI benchmark results as a dashboard
All 12 checks passedFirst working version after 8 min
Interactive explainer built by Qwen3.8 27BOpen app ↗
7 · Interactive explainer
How a language model writes, one word at a time
All 15 checks passedFirst working version after 12 min
Bird flock built by Qwen3.8 27BOpen app ↗
8 · Bird flock
A living flocking simulation you can play with
All 14 checks passedFirst working version after 25 min
Invoice generator built by Qwen3.8 27BOpen app ↗
9 · Invoice generator
An invoice that is ready to print
All 13 checks passedFirst working version after 18 min
Photo editor built by Qwen3.8 27BOpen app ↗
10 · Photo editor
A photo editor in the browser
All 14 checks passedFirst working version after 20 min
32 GB class · contender 1

Qwen3.8 27B on RTX 5090, 32 GB

llama.cpp · Unsloth UD-IQ4_XS · 128K context · KV cache q4_0 · MTP 3

10 / 10apps passed every automatic check
4 minmedian until the first working version
45 minper app, alone, no human help
Model fit calculator built by Qwen3.8 27BOpen app ↗
1 · Model fit calculator
Will this AI model fit on my graphics card?
All 16 checks passedFirst working version after 1 min
3D scene built by Qwen3.8 27BOpen app ↗
2 · 3D scene
The home AI lab at night, in real-time 3D
All 13 checks passedFirst working version after 2 min
Product page built by Qwen3.8 27BOpen app ↗
3 · Product page
A landing page for an invented home AI box
All 17 checks passedFirst working version after 3 min
Platformer game built by Qwen3.8 27BOpen app ↗
4 · Platformer game
A 2D jump and run with 3 levels
All 18 checks passedFirst working version after 5 min
Time tracker built by Qwen3.8 27BOpen app ↗
5 · Time tracker
Time tracking with project billing
All 15 checks passedFirst working version after 6 min
Dashboard built by Qwen3.8 27BOpen app ↗
6 · Dashboard
Local AI benchmark results as a dashboard
All 12 checks passedFirst working version after 4 min
Interactive explainer built by Qwen3.8 27BOpen app ↗
7 · Interactive explainer
How a language model writes, one word at a time
All 15 checks passedFirst working version after 3 min
Bird flock built by Qwen3.8 27BOpen app ↗
8 · Bird flock
A living flocking simulation you can play with
All 14 checks passedFirst working version after 7 min
Invoice generator built by Qwen3.8 27BOpen app ↗
9 · Invoice generator
An invoice that is ready to print
All 13 checks passedFirst working version after 5 min
Photo editor built by Qwen3.8 27BOpen app ↗
10 · Photo editor
A photo editor in the browser
All 14 checks passedFirst working version after 7 min
128 GB class · contender 1

Qwen3.8 Flash-Next on 1× GB10 box (Lenovo ThinkStation PGX), 128 GB

vLLM · NVFP4 · 262K context · MTP 3

Runs on an older version of @MiaAI_lab's single-box kit (commit 6b50864, 19 Sep), 13 commits behind her main at run time. I changed two things: KV cache BF16 instead of her default FP8, and CHAT_TEMPLATE pointed to the template file in her repo (her default leaves it empty). So this is not her stock recipe, and her current version may be faster.

10 / 10apps passed every automatic check
7 minmedian until the first working version
45 minper app, alone, no human help
Model fit calculator built by Qwen3.8 Flash-NextOpen app ↗
1 · Model fit calculator
Will this AI model fit on my graphics card?
All 16 checks passedFirst working version after 4 min
3D scene built by Qwen3.8 Flash-NextOpen app ↗
2 · 3D scene
The home AI lab at night, in real-time 3D
All 13 checks passedFirst working version after 22 min
Product page built by Qwen3.8 Flash-NextOpen app ↗
3 · Product page
A landing page for an invented home AI box
All 17 checks passedFirst working version after 7 min
Platformer game built by Qwen3.8 Flash-NextOpen app ↗
4 · Platformer game
A 2D jump and run with 3 levels
All 18 checks passedFirst working version after 8 min
Time tracker built by Qwen3.8 Flash-NextOpen app ↗
5 · Time tracker
Time tracking with project billing
All 15 checks passedFirst working version after 7 min
Dashboard built by Qwen3.8 Flash-NextOpen app ↗
6 · Dashboard
Local AI benchmark results as a dashboard
All 12 checks passedFirst working version after 4 min
Interactive explainer built by Qwen3.8 Flash-NextOpen app ↗
7 · Interactive explainer
How a language model writes, one word at a time
All 15 checks passedFirst working version after 4 min
Bird flock built by Qwen3.8 Flash-NextOpen app ↗
8 · Bird flock
A living flocking simulation you can play with
All 14 checks passedFirst working version after 7 min
Invoice generator built by Qwen3.8 Flash-NextOpen app ↗
9 · Invoice generator
An invoice that is ready to print
All 13 checks passedFirst working version after 14 min
Photo editor built by Qwen3.8 Flash-NextOpen app ↗
10 · Photo editor
A photo editor in the browser
All 14 checks passedFirst working version after 7 min
128 GB class · contender 2

Qwen3.6 35B-A3B on 1× GB10 box (Gigabyte AI TOP ATOM), 128 GB

vLLM · Unsloth NVFP4 · 128K context · no speculative decoding

The box froze several times during the night. I moved apps 8 and 10 to two other boxes of the same type with the same setup, so times are not shown.

9 / 10apps passed every automatic check
45 minper app, alone, no human help
Model fit calculator built by Qwen3.6 35B-A3BOpen app ↗
1 · Model fit calculator
Will this AI model fit on my graphics card?
All 16 checks passed
3D scene built by Qwen3.6 35B-A3BOpen app ↗
2 · 3D scene
The home AI lab at night, in real-time 3D
All 13 checks passed
Product page built by Qwen3.6 35B-A3BOpen app ↗
3 · Product page
A landing page for an invented home AI box
All 17 checks passed
Platformer game built by Qwen3.6 35B-A3BOpen app ↗
4 · Platformer game
A 2D jump and run with 3 levels
17 of 18 checks
Time tracker built by Qwen3.6 35B-A3BOpen app ↗
5 · Time tracker
Time tracking with project billing
All 15 checks passed
Dashboard built by Qwen3.6 35B-A3BOpen app ↗
6 · Dashboard
Local AI benchmark results as a dashboard
All 12 checks passed
Interactive explainer built by Qwen3.6 35B-A3BOpen app ↗
7 · Interactive explainer
How a language model writes, one word at a time
All 15 checks passed
Bird flock built by Qwen3.6 35B-A3BOpen app ↗
8 · Bird flock
A living flocking simulation you can play with
All 14 checks passed
Invoice generator built by Qwen3.6 35B-A3BOpen app ↗
9 · Invoice generator
An invoice that is ready to print
All 13 checks passed
Photo editor built by Qwen3.6 35B-A3BOpen app ↗
10 · Photo editor
A photo editor in the browser
All 14 checks passed
256 GB class · contender 1

GLM-5.3-Flash on 2× GB10 boxes (ASUS Ascent GX10), 256 GB

vLLM · EXL3 4bpw · tensor parallel 2 · MTP 2 · 128K context

Runs on an older version of @MiaAI_lab's 2-box kit (c1b7d4c, 22 Sep) with my own changes: MTP 2 instead of DFlash2, KV cache fixed at 3 GiB, max 2 sequences. Not her current default.
Apps 3 and 4 were never built. The engine connection dropped on every request and the model returned no tokens. That is my setup failing, not GLM or Mia's kit. I will rerun both. 8 of 8 completed apps passed every check.

8 / 10apps passed every automatic check
18 minmedian until the first working version
45 minper app, alone, no human help
Model fit calculator built by GLM-5.3-FlashOpen app ↗
1 · Model fit calculator
Will this AI model fit on my graphics card?
All 16 checks passedFirst working version after 10 min
3D scene built by GLM-5.3-FlashOpen app ↗
2 · 3D scene
The home AI lab at night, in real-time 3D
All 13 checks passedFirst working version after 6 min
Product page: never built
3 · Product page
A landing page for an invented home AI box
Never built (my engine connection dropped)
Platformer game: never built
4 · Platformer game
A 2D jump and run with 3 levels
Never built (my engine connection dropped)
Time tracker built by GLM-5.3-FlashOpen app ↗
5 · Time tracker
Time tracking with project billing
All 15 checks passedFirst working version after 42 min
Dashboard built by GLM-5.3-FlashOpen app ↗
6 · Dashboard
Local AI benchmark results as a dashboard
All 12 checks passedFirst working version after 18 min
Interactive explainer built by GLM-5.3-FlashOpen app ↗
7 · Interactive explainer
How a language model writes, one word at a time
All 15 checks passedFirst working version after 18 min
Bird flock built by GLM-5.3-FlashOpen app ↗
8 · Bird flock
A living flocking simulation you can play with
All 14 checks passedFirst working version after 20 min
Invoice generator built by GLM-5.3-FlashOpen app ↗
9 · Invoice generator
An invoice that is ready to print
All 13 checks passedFirst working version after 15 min
Photo editor built by GLM-5.3-FlashOpen app ↗
10 · Photo editor
A photo editor in the browser
All 14 checks passedFirst working version after 22 min
Side by side

One prompt, every model

Pick a task and see what each machine made of the same brief (Round 1).

Overview

Task × setup, Round 1

Cell: automatic checks passed (45 min per app, one run per task).

Qwen3.8 27B
RTX 5060 Ti		Qwen3.8 27B
RTX 5090		Qwen3.8 Flash-Next
1× GB10 box (Lenovo ThinkStation PGX)		Qwen3.6 35B-A3B
1× GB10 box (Gigabyte AI TOP ATOM)		GLM-5.3-Flash
2× GB10 boxes (ASUS Ascent GX10)
1 · Model fit calculator16/1616/1616/1616/1616/16
2 · 3D scene13/1313/1313/1313/1313/13
3 · Product page17/1717/1717/1717/17–
4 · Platformer game18/1818/1818/1817/18–
5 · Time tracker15/1515/1515/1515/1515/15
6 · Dashboard12/1212/1212/1212/1212/12
7 · Interactive explainer15/1515/1515/1515/1515/15
8 · Bird flock14/1414/1414/1414/1414/14
9 · Invoice generator13/1313/1313/1313/1313/13
10 · Photo editor14/1414/1414/1414/1414/14
How it was measured

Same brief, same tools, same clock

  1. One brief per app. The model gets the prompt, a shell and a real browser. Nobody answers questions.
  2. 45 minutes, then stop (Round 2: 45 or 150 depending on the task). The model commits its work as it goes. The version that counts is its last commit that passes the checks.
  3. Automatic checks. A script opens every app and clicks through it: buttons, keyboard, totals, level changes, dark mode and more. Round 2 scores tiers: must-pass, quality and top checks.
  4. Everything runs at home. Models run on my own machines, the agent is OpenCode. No cloud model was used to build any app.

The benchmark runs, the checks and this page were set up by AI agents working for me. The apps on this page were written entirely by the local models named above.