Qwen, Ministral, Gemma and Llama: local model comparison

This comparison uses synthesized narration. This local LLM comparison ranks nine models against a hypothetical checkout bug: inspect a screenshot, trace an incorrect total through code, and explain the fix to a client. The speaker judges capability alongside memory use and responsiveness. The tiers reflect that workload, rather than a universal ranking.

Qwen 3.8 27B takes the speaker's sole S tier for image input, coding results and adjustable reasoning effort. Its four-bit package is about 19GB, plus roughly 1GB for a separate vision component. The suggested working budget is 22–26GB of VRAM. A 24GB card is described as plausible with controlled context; partial offloading on a 12GB card may sacrifice responsiveness.

Ministral 3 8B Instruct, release 2512, earns A tier as a compact visual assistant, with an estimated 8–10GB VRAM budget. Gemma 4 26B A4B also earns A, but its sparse computation does not remove the need to store the full model. Devstral Small 2 24B earns A for repository repair, with a suggested 18–24GB budget.

The text-only Llama 3.3 70B receives D for this workflow despite its size: the download is about 43GB. Phi 4 Reasoning Vision 15B receives B, with runner compatibility and its 16,384-token context limiting broader use as a coding assistant. All memory ranges are planning estimates for one user and modest context. The speaker also distinguishes vendor benchmark results and AI agent setups from guarantees for compressed local downloads.