Alex Ellis argues a local Qwen is a different tool from Opus, not a worse one

Category: Agentic coding
Status: discussion
Sources: blog.alexellis.io, discussion
Summary: Alex Ellis published a write-up on 2026-06-18 on running local Qwen coding models on owned hardware: an RTX 6000 Pro Blackwell with 96GB VRAM plus two RTX 3090s, two llama.cpp instances at the full 262,144-token context with an F16 KV cache, generating about 67 tokens per second and 130 to 200 with speculative decoding. His thesis is that a local 27B Qwen is not a cheaper near-Opus but a different tool, strong on bounded tasks such as customer diagnostics, telemetry and revenue-anomaly review, and codebase explanation, and weak on long-horizon unsupervised work, concurrency, and some Go generation. He calls infinite looping the model's worst failure mode, where it repeats suggestions or stalls at a capability boundary and needs manual supervision. The post cites Qwen at 77.2 percent on SWE-Bench Verified against 88.6 percent for Claude Opus 4.8; the figures are as stated in the post, not independently reproduced.
Why it matters: It is a concrete hardware-and-numbers account of where local coding models fit in real workflows, a counterweight to near-Opus framing as open-weight models such as GLM-5.2 and Qwen proliferate.