Hybrid attention speeds Rust code model 50x

Strong research signal, but this reads more like a systems and scaling note than a product launch.

• –The clearest result is that data scaling beat architecture tuning at this size; that is the most defensible takeaway.
• –The inference win is substantial, but it appears to come from the cache/compression strategy as much as from the attention formulation itself.
• –Quality evidence is still thin: perplexity is good for a tiny model, but code usefulness should be judged with parsing, compilation, and completion benchmarks.
• –The post would be stronger with ablations for hybrid vs local-only vs recurrent-only, plus earlier-checkpoint generation samples.
• –For a model this small, longer context and better tokenization are likely to matter as much as more exotic attention variants.