MULTI-AGENT-LAB
Test models.
Give them clear jobs.
Which local model helps with which task? In the lab we compare answers, inspect failures and test how the parts work together in Folio.
01 / CONFIGURATION · SEPTEMBER 2026
A team with clear roles.
Daily mail assessment combines fixed rules with three model voices. Disagreement can trigger an additional review.
GLM-4.7 Flash
A first control assessment for classification.
Qwen 3.6
A second control assessment alongside the primary one.
Qwen 3.8
Primary assessment and model for the local companion.
Gemma 4 31B
An extra check when assessments disagree.
02 / THE MODEL CATALOG
Twelve models.
Different experiments.
Models for mail classification, knowledge selection and task interpretation.
All model names and variants
| Model | Local variant |
|---|---|
| Qwen 3.6 35B A3B | MLX · 4-bit |
| Gemma 4 31B QAT | GGUF · Q4_0 |
| Qwen 3.8 27B | MLX · 4-bit |
| Gemma 4 26B A4B | MLX · 4-bit |
| Qwen 3 30B A3B Thinking | MLX · 4-bit |
| Ministral 3 14B Instruct | GGUF · Q4_K_M |
| NVIDIA Nemotron 3 Nano 30B A3B | GGUF · Q4_K_M |
| Liquid LFM2 24B A2B | GGUF · Q6_K |
| IBM Granite 4.1 8B | GGUF · Q8_0 |
| GLM-4.7 Flash 30B A3B | MLX · 6-bit |
| Phi-4 Mini Instruct | GGUF · Q8_0 |
| Muse Glimmer 30B | GGUF · Q4_K_M · archive |
03 / FROM THE TESTS
Small tests.
Close inspection.
Which model spots a task reliably — and how quickly? A comparison using fourteen short messages.
24.09.2026 · 14 FICTIONAL EXAMPLES
What actually counts as a task?
Three models classify the same messages: task, unclear or no task. Standard instruction (v1).
Speed matters.
Gemma 4 26B was fast and accurate on these short tasks.
The whole path matters.
For Ledger we test language understanding and tools separately. The model interprets the question; code queries transactions and calculates totals.
Choose for the job.
Mail sorting and financial questions need different skills. We test models on the tasks they will handle.