Qwen3.5 4B, 35B pair well locally

This is the kind of practical local-model workflow insight that matters more than leaderboard bragging rights: small models can add value as editors, critics, and sanity-checkers instead of trying to beat bigger models head-on.

• –The core takeaway is operational, not academic: Qwen3.5 4B is fast enough to be useful in an iterative loop, while 35B is slow but still usable on consumer hardware with tuning
• –That makes a two-model setup plausible for local power users who want one model for drafting speed and another for broader coverage
• –It also reinforces a growing open-model pattern: smaller models are increasingly good at review, extraction, and comparison tasks even when they are not the best primary generators
• –Because this is a single-user qualitative test, developers should treat it as an interesting workflow pattern rather than a definitive benchmark result
• –The post is most relevant for people building local inference stacks around consumer GPUs, Jan, and quantized open-weight models