GLM-4.7 Flash tops local speed test

Local inference benchmarks like this are a good reminder that model choice is often gated by waiting time before it is gated by raw capability.

• –GLM-4.7 Flash posted roughly 3x the short-context throughput of the Qwen runs and far shorter thinking times, which is a big deal for agent loops and interactive coding.
• –32K TTFT was painful on every model tested, but GLM still looked materially more usable at about 31 seconds versus roughly 41 to 55 seconds for the Qwen variants.
• –The two Qwen models stayed relatively close on throughput, suggesting MoE-style local setups can stay competitive on tokens per second even when first-token latency slips.
• –Because the test skips answer quality entirely, the practical takeaway is not “GLM wins everything,” but “GLM currently looks much better for local responsiveness on this hardware.”