Community Discussion · Tracks

Teams should seriously consider the BitNet direction

Kevin_GuKevin_GuAug 112026/08/11 275 views

Let me share something. Today I stumbled upon Portail, an AI gateway written in Rust, featuring a SIMD tensor engine for BitNet b1.58 and claiming a throughput of 16.9k tok/s. The numbers look impressive, but what really caught my attention was the underlying model strategy.

2 replies

?
Ctrl + Enter to reply
Jiang Shouqian

The 0.4GB figure is certainly attractive, but as a pre-sales engineer, I need to throw some cold water on it. Clients usually don't just look at memory; they care more about the actual trade-off between accuracy and throughput. Portail's SIMD engine looks good in ideal scenarios, but when running in real client businesses, long-tail scenarios and model compatibility are key. I suggest the team first stress-test with a real client scenario to see if willingness to pay can sustain it.

Mo Mo
Mo MoAug 11

0.4GB is indeed a bit absurd... I just wrote a post on ComputeNet last week about scheduling issues. If this route actually works out, the whole approach to scheduling might need to change. But then again, while the SIMD engine looks good on performance benchmarks, actual engineering implementation depends on the ecosystem. Just having a Rust gateway isn't enough, right?