https://seqpu.com/CPUsArentDead/
Article
-
Gemma 4 2B runs on CPU, matches GPT-3.5 Turbo on MT-Bench
-
Authors added surgical guardrails (calculator, logic solver, regex) to boost scores
-
Argues capable local inference on commodity hardware is now viable
Discussion
-
Commenters note guardrails are benchmark-tuned, not general capability improvements
-
Comparison to GPT-3.5 called weak; Opus 4.5/4.6 are actual current SOTA
-
Genuine excitement about Gemma 4 quality for local/offline use cases
Discuss on HN