Gemma 4, Qwen3.6 beat old 70Bs

The 70B floor is losing relevance fast; for local RP and general use, newer 26B-31B-class models are starting to win on practical quality-per-watt, not just raw size.

• –Gemma 4 is getting the strongest praise here, especially for prompt adherence, longer-session coherence, and cleaner prose than older 70B baselines.
• –Qwen’s newer releases are still competitive, but the thread skews toward Gemma for roleplay style, while Qwen looks stronger as a general-purpose or agentic model family.
• –Memory efficiency matters as much as quality: lower KV-cache demand and better context handling make these models much easier to run on consumer hardware.
• –The tradeoff is still real: some users note token duplication, occasional slop, and the fact that finetunes are still catching up.
• –Net: if you can’t run 70B, that’s less of a handicap than it used to be.