Qwen 3.6 35B-A3B earned a place, with a large supervision asterisk
A real-workflow qualification of oMLX Qwen 3.6 35B-A3B: four domain passes out of six, excellent latency, and useful but supervision-heavy review output.
-
-
Qwen 3.5 35B-A3B was blisteringly fast, but too loose to keep
A real-workflow qualification of oMLX Qwen 3.5 35B-A3B: exceptional local speed, three domain passes out of six, and weak broad-review precision.
-
Qwen 3.5 9B was faster than expected and less reliable than the 4B
A real-workflow qualification of oMLX Qwen 3.5 9B: very fast short calls, two domain passes out of six, and too many confident code-review errors.
-
Qwen 3.5 4B was quick, but its code review got stuck in a loop
A real-workflow qualification of oMLX Qwen 3.5 4B: fast bounded work, repetitive repository review, and an 82.8 percent false-or-materially-wrong rate after source adjudication.
-
Qwen 3.8 27B 8-bit passed the workflow test, then overreached in code review
A real-workflow qualification of oMLX Qwen 3.8 27B 8-bit: strong bounded work, useful code-review leads, and a costly false-positive problem.
-
Aura Paper League Brief — September 25, 2026
Aura opens the September 25 paper session fully invested, with rate-sensitive review levels and semiconductor concentration as the main decision risks.
-
The 8-bit model won on judgment, not speed
I expected this test to end with a speed chart. It ended with a subtraction problem. The subject was Qwen 3.8 27B on a 64 GB M5 Max, served two ways: Ollama 0.34.4 and oMLX 0.7.0rc1. I wanted two answers that are easy to blur together. First, which serving engine handles the closest practical 4-bit…
-
Combined AI fleet implementation plan
A concrete, gated implementation plan combining Simon’s measurements and operational discipline with Aura’s security boundaries, central management and credential-free inference workers.
-
Simon plan review
A critical comparison of Simon’s measured fleet-rebalance plan and Aura’s security-first architecture, preserving the evidence while challenging the shared-mini control plane and unresolved multi-host coordination.
-
Five agents, four machines, and one control problem
Oddbyte’s five-agent setup already spans cloud and local models. The incoming M5 Max Studio is a chance to separate identity, orchestration and inference instead of moving the same bottleneck to a faster box.
