Gemma 3 27B sparks scriptwriting hunt

This is less about benchmark bragging and more about finding a model that can think like a director. Gemma 3 is the closest open-weight Google family member, but it should be treated as Gemini's cousin, not a clone of the closed stack.

• –Google says Gemma 3 shares research and technology lineage with Gemini 2.0 and uses the same tokenizer, but the model card describes its own 14T-token training mix built from web text and code.
• –The 27B tier plus official quantized versions make it the most plausible fit for a serious local setup, especially if the target is a single-GPU budget.
• –The 128K context window matters here because script outlines, beat sheets, and character notes all benefit from long-form composition in one prompt.
• –Vision support is nice but not the core ask; the real differentiator is whether the model can maintain structure, timing, and emotional cues across a scene.
• –The thread exposes a real gap in local LLMs: many models can describe a shot, but far fewer can direct one.