LocalLLaMA debates May model wave

The thread reads like a realistic temperature check on the local-LLM market: more of the same from the frontier vendors is likely, while the big unknown is whether any release meaningfully changes inference cost, quantization quality, or coder utility.

• –Most plausible winners are incremental expansions from Gemma, Qwen, Mistral, DeepSeek, and GLM, because those families already have momentum in local deployment
• –Bigger models are less exciting than better small and mid-size variants, since local users care more about latency, memory footprint, and quantized quality
• –A true surprise would come from a hardware player or an OpenAI OSS drop that is actually practical for local use, not just a research showcase
• –The most useful advances may be method-level: better distillation, stronger reasoning at smaller sizes, cleaner MoE routing, and fewer quantization regressions
• –The wishlist is telling: developers want models that are easier to run, easier to tune, and easier to integrate into agentic workflows