A cyber bench is a quoted speed and a name. A lab is not a speed. A dash means they printed none.
A cyber bench ranks a published decode, with the publisher's name on the row. The first shelf is Cyber. This desk did not remeasure it.
What gets a rank
Only a decode the publisher printed. Highest first. CYBER-FROST 3.8 is one weight and three serving modes, because the card printed three. BF16 with MTP k=2 is 265.00 tok/s on 4× B300. MTP k=3 is 263.10. No draft is 190.10. The NVFP4 file on one DGX Spark is 29.29 tok/s on code and 26.40 on prose. FP8 has a dash. Those rows are Cyber. The weight itself is Model.8. The files are huggingface.co/Blackfrost-AI/CYBER-FROST-3.8-BF16.
Whose name
The cyber weights are Blackfrost's, x.com/Blackfrost_AI. The foundation is Qwen3.8-Flash-Next. Blackfrost thanked x.com/QwenDevs for that base. Their card says the weight is for authorized security research, assessment, and response. It also says it does not authorize access, choose a target, or replace a legal review. This desk does not turn that card into a how-to.
What does not get a rank
bugbounty-lab101 and T3MP3ST belong to DevCop95, github.com/DevCop95/bugbounty-lab101. They sit under the table. A workspace and an agent room are not a tok/s. The GGUF at huggingface.co/freakyskittle/CYBER-FROST-3.8-GGUF is not ranked. Blackfrost pointed at the file. The author says no quant there has passed a smoke test. A star count is not a bench.
What a later bench needs
The same prompt, the same box, the same context, and a number this desk can reproduce, still with the publisher's name on the row. Until then the row says quoted, not measured. Hardware stays the street serving board. Payback stays a US street-price clock. A 4× B300 trial does not belong in that clock. Mixing them would be a lie.