Local LLM Choice

Repairing local model artifacts

A clean comparison of each local model's first idea, its single-shot self-repair, bounded local repair loops, a cross-model Qwen3-Coder repair pass, and a GPT-5.5 upper-bound implementation.

5Models
5Base runs
5Fix runs
11Agentic attempts
5Pi agent runs
4Qwen-Coder repairs
5GPT-5.5 runs
14Smoke passes

This publication-facing view groups artifacts by model and keeps the progression readable without runner operational noise.

Qwen

qwen3-coder-30b

mlx-community/Qwen3-Coder-30B-A3B-Instruct-4bit
Base

Interactive Visualization of the Fibonacci Sequence in Nature

2026-07-05-qwen3-coder-30b-006
FailedBrokenSmoke failed
Idea 1
Exec 1
Taste 1

Browser console shows ReferenceError: ffibSequence is not defined on initial load and again when the range input fires.

Fix

Interactive Fibonacci Spiral Visualizer with Natural Patterns

2026-07-05-qwen3-coder-30b-self-repair-001
PartialNot applicableSmoke not applicable
Idea 2
Exec 2
Taste 2

Artifact is not HTML; smoke test skipped.

Attempts

Fibonacci Spiral Visualizer with Interactive Square Labels

2026-07-05-qwen3-coder-30b-agentic-repair-003
FailedBrokenSmoke failed
Idea 1
Exec 1
Taste 1

Attempt 1: Failed; Attempt 2: Failed; Attempt 3: Failed

Agentic Loop

Fibonacci Spiral Visualizer with Interactive Square Labels

2026-07-05-qwen3-coder-30b-agentic-repair-003
FailedBrokenSmoke failed
Idea 1
Exec 1
Taste 1

Runtime smoke error during input event fibSlider: ReferenceError: ffibSequence is not defined

Pi Agent

Interactive Visualization of the Fibonacci Sequence in Nature

2026-07-06-qwen3-coder-30b-pi-agent-001
GoodWorksSmoke passed
Idea 3
Exec 4
Taste 3

Local pi-coding-agent repair against the copied first-round artifact.

Qwen-Coder Repair

Not run

Pending

Cross-model local repair track; Qwen3-Coder is used as the repair agent.

GPT-5.5 Version

Interactive Visualization of the Fibonacci Sequence in Nature

2026-07-06-qwen3-coder-30b-gpt55-version-001
GoodWorksSmoke passed
Idea 3
Exec 4
Taste 3

Local smoke test passed: no external URLs, syntax failures, or smoke-harness runtime errors were found.

Qwen

qwen3-30b

mlx-community/Qwen3-30B-A3B-4bit
Base

Prime Spiral Visualizer

2026-07-05-qwen3-30b-001
GoodWorksSmoke passed
Idea 3
Exec 3
Taste 2

Browser inspection showed a visible 800x800 canvas with a rendered prime pattern and no qwen3-30b console errors.

Fix

Prime Spiral Enhanced

2026-07-05-qwen3-30b-self-repair-001
PartialNot applicableSmoke not applicable
Idea 2
Exec 2
Taste 2

Artifact is not HTML; smoke test skipped.

Attempts

Prime Spiral Visualizer (Enhanced)

2026-07-05-qwen3-30b-agentic-repair-001
GoodWorksSmoke passed
Idea 3
Exec 4
Taste 3

Attempt 1: Good

Agentic Loop

Prime Spiral Visualizer (Enhanced)

2026-07-05-qwen3-30b-agentic-repair-001
GoodWorksSmoke passed
Idea 3
Exec 4
Taste 3

Local smoke test passed: no external URLs, syntax failures, or smoke-harness runtime errors were found.

Pi Agent

Prime Spiral Visualizer

2026-07-06-qwen3-30b-pi-agent-001
FailedBrokenSmoke failed
Idea 1
Exec 1
Taste 1

Local pi-coding-agent repair against the copied first-round artifact.

Qwen-Coder Repair

Prime Spiral Visualizer

2026-07-06-qwen3-30b-qwen-coder-pi-repair-001
GoodWorksSmoke passed
Idea 3
Exec 4
Taste 3

Qwen3-Coder repairs this model's copied first-round artifact through Pi.

GPT-5.5 Version

Prime Spiral Visualizer

2026-07-06-qwen3-30b-gpt55-version-001
GoodWorksSmoke passed
Idea 3
Exec 4
Taste 3

Local smoke test passed: no external URLs, syntax failures, or smoke-harness runtime errors were found.

Gemma

gemma3-27b

mlx-community/gemma-3-text-27b-it-4bit
Base

Italian Phrasebook Generator

2026-07-05-gemma3-27b-001
PartialWorksSmoke passed
Idea 1
Exec 3
Taste 2

Browser testing found the Generate Phrase button and confirmed it changed the display to an Italian-English phrase.

Fix

Italian Phrasebook with Themed Categories

2026-07-05-gemma3-27b-self-repair-001
GoodWorksSmoke passed
Idea 3
Exec 3
Taste 3

Local smoke test passed: no external URLs, syntax failures, or smoke-harness runtime errors were found.

Attempts

Italian Phrasebook with Context and Difficulty

2026-07-05-gemma3-27b-agentic-repair-001
GoodWorksSmoke passed
Idea 3
Exec 4
Taste 3

Attempt 1: Good

Agentic Loop

Italian Phrasebook with Context and Difficulty

2026-07-05-gemma3-27b-agentic-repair-001
GoodWorksSmoke passed
Idea 3
Exec 4
Taste 3

Local smoke test passed: no external URLs, syntax failures, or smoke-harness runtime errors were found.

Pi Agent

Italian Phrasebook Generator

2026-07-06-gemma3-27b-pi-agent-001
GoodWorksSmoke passed
Idea 3
Exec 4
Taste 3

Local pi-coding-agent repair against the copied first-round artifact.

Qwen-Coder Repair

Italian Phrasebook Generator

2026-07-06-gemma3-27b-qwen-coder-pi-repair-001
GoodWorksSmoke passed
Idea 3
Exec 4
Taste 3

Qwen3-Coder repairs this model's copied first-round artifact through Pi.

GPT-5.5 Version

Italian Phrasebook Generator

2026-07-06-gemma3-27b-gpt55-version-001
GoodWorksSmoke passed
Idea 3
Exec 4
Taste 3

Local smoke test passed: no external URLs, syntax failures, or smoke-harness runtime errors were found.

Mistral

devstral-24b

mlx-community/Devstral-Small-2505-4bit
Base

Interactive Data Visualization of Global Earthquakes

2026-07-05-devstral-24b-001
PartialPartially worksSmoke failed
Idea 2
Exec 1
Taste 2

Source uses external Leaflet from unpkg.com, OpenStreetMap tiles, and a live USGS fetch, so it is not a self-contained local artifact.

Fix

Interactive Data Visualization of Global Earthquakes

2026-07-05-devstral-24b-self-repair-001
FailedBrokenSmoke failed
Idea 1
Exec 1
Taste 1

Artifact references external URLs or remote fetches: https://unpkg.com/leaflet@1.7.1/dist/leaflet.js

Attempts

Interactive Visualization of Global Earthquakes (Self-Contained)

2026-07-05-devstral-24b-agentic-repair-003
FailedBrokenSmoke failed
Idea 1
Exec 1
Taste 1

Attempt 1: Failed; Attempt 2: Failed; Attempt 3: Failed

Agentic Loop

Interactive Visualization of Global Earthquakes (Self-Contained)

2026-07-05-devstral-24b-agentic-repair-003
FailedBrokenSmoke failed
Idea 1
Exec 1
Taste 1

Controls were present, but the smoke harness saw no DOM, canvas, alert, or child-list changes.

Pi Agent

Interactive Data Visualization of Global Earthquakes

2026-07-06-devstral-24b-pi-agent-001
FailedBrokenSmoke failed
Idea 1
Exec 1
Taste 1

Local pi-coding-agent repair against the copied first-round artifact.

Qwen-Coder Repair

Interactive Data Visualization of Global Earthquakes

2026-07-06-devstral-24b-qwen-coder-pi-repair-001
FailedBrokenSmoke failed
Idea 1
Exec 1
Taste 1

Qwen3-Coder repairs this model's copied first-round artifact through Pi.

GPT-5.5 Version

Interactive Data Visualization of Global Earthquakes

2026-07-06-devstral-24b-gpt55-version-001
GoodWorksSmoke passed
Idea 3
Exec 4
Taste 3

Local smoke test passed: no external URLs, syntax failures, or smoke-harness runtime errors were found.

DeepSeek

deepseek-r1-14b

mlx-community/DeepSeek-R1-Distill-Qwen-14B-4bit
Base

Interactive Data Visualization Tool

2026-07-05-deepseek-r1-14b-001
FailedBrokenSmoke failed
Idea 1
Exec 1
Taste 1

Browser testing showed the Generate button exists, but clicking it logs ReferenceError: canvas is not defined from drawChart().

Fix

Self Repair: Interactive Data Visualization Tool

2026-07-05-deepseek-r1-14b-self-repair-001
FailedBrokenSmoke failed
Idea 1
Exec 1
Taste 1

Runtime smoke error during onclick button: ReferenceError: chartStartX is not defined

Attempts

Agentic Repair: Interactive Data Visualization Tool

2026-07-05-deepseek-r1-14b-agentic-repair-003
FailedBrokenSmoke failed
Idea 1
Exec 1
Taste 1

Attempt 1: Failed; Attempt 2: Failed; Attempt 3: Failed

Agentic Loop

Agentic Repair: Interactive Data Visualization Tool

2026-07-05-deepseek-r1-14b-agentic-repair-003
FailedBrokenSmoke failed
Idea 1
Exec 1
Taste 1

Runtime smoke error during onclick button: ReferenceError: canvas is not defined

Pi Agent

Interactive Data Visualization Tool

2026-07-06-deepseek-r1-14b-pi-agent-001
FailedBrokenSmoke failed
Idea 1
Exec 1
Taste 1

Local pi-coding-agent repair against the copied first-round artifact.

Qwen-Coder Repair

Interactive Data Visualization Tool

2026-07-06-deepseek-r1-14b-qwen-coder-pi-repair-001
FailedBrokenSmoke failed
Idea 1
Exec 1
Taste 1

Qwen3-Coder repairs this model's copied first-round artifact through Pi.

GPT-5.5 Version

Interactive Data Visualization Tool

2026-07-06-deepseek-r1-14b-gpt55-version-001
GoodWorksSmoke passed
Idea 3
Exec 4
Taste 3

Local smoke test passed: no external URLs, syntax failures, or smoke-harness runtime errors were found.

Run Index

ModelStageRunStatusLinks
deepseek-r1-14b
Base
Interactive Data Visualization Tool
2026-07-05-deepseek-r1-14b-001
FailedBrokenSmoke failed
artifact report evaluation
deepseek-r1-14b
Agentic 1
Interactive Data Visualization Tool (Fixed)
2026-07-05-deepseek-r1-14b-agentic-repair-001
FailedBrokenSmoke failed
artifact report evaluation
deepseek-r1-14b
Agentic 2
Agentic Repair: Interactive Data Visualization Tool
2026-07-05-deepseek-r1-14b-agentic-repair-002
FailedBrokenSmoke failed
artifact report evaluation
deepseek-r1-14b
Agentic 3
Agentic Repair: Interactive Data Visualization Tool
2026-07-05-deepseek-r1-14b-agentic-repair-003
FailedBrokenSmoke failed
artifact report evaluation
deepseek-r1-14b
Fix 1
Self Repair: Interactive Data Visualization Tool
2026-07-05-deepseek-r1-14b-self-repair-001
FailedBrokenSmoke failed
artifact report evaluation
devstral-24b
Base
Interactive Data Visualization of Global Earthquakes
2026-07-05-devstral-24b-001
PartialPartially worksSmoke failed
artifact report evaluation
devstral-24b
Agentic 1
Interactive Data Visualization of Global Earthquakes - Self-Contained
2026-07-05-devstral-24b-agentic-repair-001
FailedBrokenSmoke failed
artifact report evaluation
devstral-24b
Agentic 2
Interactive Data Visualization of Global Earthquakes
2026-07-05-devstral-24b-agentic-repair-002
FailedBrokenSmoke failed
artifact report evaluation
devstral-24b
Agentic 3
Interactive Visualization of Global Earthquakes (Self-Contained)
2026-07-05-devstral-24b-agentic-repair-003
FailedBrokenSmoke failed
artifact report evaluation
devstral-24b
Fix 1
Interactive Data Visualization of Global Earthquakes
2026-07-05-devstral-24b-self-repair-001
FailedBrokenSmoke failed
artifact report evaluation
gemma3-27b
Base
Italian Phrasebook Generator
2026-07-05-gemma3-27b-001
PartialWorksSmoke passed
artifact report evaluation
gemma3-27b
Agentic 1
Italian Phrasebook with Context and Difficulty
2026-07-05-gemma3-27b-agentic-repair-001
GoodWorksSmoke passed
artifact report evaluation
gemma3-27b
Fix 1
Italian Phrasebook with Themed Categories
2026-07-05-gemma3-27b-self-repair-001
GoodWorksSmoke passed
artifact report evaluation
qwen3-30b
Base
Prime Spiral Visualizer
2026-07-05-qwen3-30b-001
GoodWorksSmoke passed
artifact report evaluation
qwen3-30b
Agentic 1
Prime Spiral Visualizer (Enhanced)
2026-07-05-qwen3-30b-agentic-repair-001
GoodWorksSmoke passed
artifact report evaluation
qwen3-30b
Fix 1
Prime Spiral Enhanced
2026-07-05-qwen3-30b-self-repair-001
PartialNot applicableSmoke not applicable
artifact report evaluation
qwen3-coder-30b
Base
Interactive Visualization of the Fibonacci Sequence in Nature
2026-07-05-qwen3-coder-30b-006
FailedBrokenSmoke failed
artifact report evaluation
qwen3-coder-30b
Agentic 1
Fibonacci Spiral in Nature: Interactive Visualization
2026-07-05-qwen3-coder-30b-agentic-repair-001
FailedBrokenSmoke failed
artifact report evaluation
qwen3-coder-30b
Agentic 2
F Fibonacci Spiral Visualizer with Interactive Fibonacci Explorer
2026-07-05-qwen3-coder-30b-agentic-repair-002
FailedBrokenSmoke failed
artifact report evaluation
qwen3-coder-30b
Agentic 3
Fibonacci Spiral Visualizer with Interactive Square Labels
2026-07-05-qwen3-coder-30b-agentic-repair-003
FailedBrokenSmoke failed
artifact report evaluation
qwen3-coder-30b
Fix 1
Interactive Fibonacci Spiral Visualizer with Natural Patterns
2026-07-05-qwen3-coder-30b-self-repair-001
PartialNot applicableSmoke not applicable
artifact report evaluation
deepseek-r1-14b
GPT-5.5 1
Interactive Data Visualization Tool
2026-07-06-deepseek-r1-14b-gpt55-version-001
GoodWorksSmoke passed
artifact report evaluation
deepseek-r1-14b
Pi Agent 1
Interactive Data Visualization Tool
2026-07-06-deepseek-r1-14b-pi-agent-001
FailedBrokenSmoke failed
artifact report evaluation
deepseek-r1-14b
Qwen-Coder Repair 1
Interactive Data Visualization Tool
2026-07-06-deepseek-r1-14b-qwen-coder-pi-repair-001
FailedBrokenSmoke failed
artifact report evaluation
devstral-24b
GPT-5.5 1
Interactive Data Visualization of Global Earthquakes
2026-07-06-devstral-24b-gpt55-version-001
GoodWorksSmoke passed
artifact report evaluation
devstral-24b
Pi Agent 1
Interactive Data Visualization of Global Earthquakes
2026-07-06-devstral-24b-pi-agent-001
FailedBrokenSmoke failed
artifact report evaluation
devstral-24b
Qwen-Coder Repair 1
Interactive Data Visualization of Global Earthquakes
2026-07-06-devstral-24b-qwen-coder-pi-repair-001
FailedBrokenSmoke failed
artifact report evaluation
gemma3-27b
GPT-5.5 1
Italian Phrasebook Generator
2026-07-06-gemma3-27b-gpt55-version-001
GoodWorksSmoke passed
artifact report evaluation
gemma3-27b
Pi Agent 1
Italian Phrasebook Generator
2026-07-06-gemma3-27b-pi-agent-001
GoodWorksSmoke passed
artifact report evaluation
gemma3-27b
Qwen-Coder Repair 1
Italian Phrasebook Generator
2026-07-06-gemma3-27b-qwen-coder-pi-repair-001
GoodWorksSmoke passed
artifact report evaluation
qwen3-30b
GPT-5.5 1
Prime Spiral Visualizer
2026-07-06-qwen3-30b-gpt55-version-001
GoodWorksSmoke passed
artifact report evaluation
qwen3-30b
Pi Agent 1
Prime Spiral Visualizer
2026-07-06-qwen3-30b-pi-agent-001
FailedBrokenSmoke failed
artifact report evaluation
qwen3-30b
Qwen-Coder Repair 1
Prime Spiral Visualizer
2026-07-06-qwen3-30b-qwen-coder-pi-repair-001
GoodWorksSmoke passed
artifact report evaluation
qwen3-coder-30b
GPT-5.5 1
Interactive Visualization of the Fibonacci Sequence in Nature
2026-07-06-qwen3-coder-30b-gpt55-version-001
GoodWorksSmoke passed
artifact report evaluation
qwen3-coder-30b
Pi Agent 1
Interactive Visualization of the Fibonacci Sequence in Nature
2026-07-06-qwen3-coder-30b-pi-agent-001
GoodWorksSmoke passed
artifact report evaluation