Reddit user says Opus 5.5 beats benchmarks but disappoints in real work

Qwen Code's v0.27.0 release fixes managed-agent title settlement when the dispatcher claims a title first, and retargets the lifecycle integration test to retry snapshot claim deadlocks. Interactive tests now poll the foreground-shell frame instead of snapshotting it.

Modal CTO Akshat Bubna walks through what AI agents need from the infrastructure underneath them, using a spreadsheet where every cell is backed by its own agent as the example. Building blocks include a workspace per agent and state that persists for days without paying for idle time.

Out-of-core MoE streaming hits 19 tok/s decode and 62 tok/s prefill on a 12GB RTX 3060 plus 16GB RAM. The post claims a fix for the prefill bottleneck that previously made prompt processing unusable on large MoEs.

Tutorial builds NVIDIA's pinned llama.cpp fork and loads the 3.7 GB GGUF for Dory 1.0 4B, a Nemotron Labs reasoning model whose shared middle block can be applied one to four times per token. It measures decode speed and GSM8K accuracy at each loop count, plus whether the cheaper last-loop cache changes predictions.