benchmark result I don't trust: anyone testing this?

[batchfield55]

Can you fill in a bit more context? What model, what runtime, and roughly what context size?

If you post your exact settings, people here can probably narrow it down pretty fast.