FastFlowLM Linux support exposes benchmark spread

FastFlowLM’s Linux support looks legit enough to run real-world local AI tests, but the numbers make the usual tradeoff obvious: model choice matters more than raw runtime hype.

• –`lfm2.5-tk:1.2b` and `lfm2.5-it:1.2b` are the short-context speed leaders, landing around 64 tok/s generation.
• –Long-context usage is the stress test, and bigger models pay for it hard; `qwen3:8b` falls from 10.3 tok/s at 0 context to 3.6 tok/s at 70k.
• –`gpt-oss:20b` is the most interesting middle ground, with solid prefill at moderate context but a steady slide as the window grows.
• –`gemma3` and `medgemma` stay comparatively stable across deeper contexts, which suggests those families are better tuned for this stack.
• –`deepseek-r1:8b` is the oddball: generation stays flat while prefill scales up sharply, suggesting a different runtime profile than the rest.