Testing out MiMo-V2.6 Pro through OpenCode go. It's very slow... mostly around 20-24 tok/s but sometimes even dips below 10 tok/s 🐢

We got 55 TPS in @CommandCodeAI 🐐 Have at it!!

MiMo-V2.6-Flash-RL collapsed decode speed problem. Possible fix Yes that had been the problem . I fixed it with claude on my pro 6000 The fix was enabling split-KV attention for MTP’s verification pass. SGLang’s FA4 backend on your SM120 GPUs forced verification to use num_splits=1, making it scan the entire KV cache with too little GPU parallelism. We patched verification on full-attention layers to use the decode split policy—automatically allowing up to 128 splits.

@opencode interesting

@opencode It's very slow. Pretty much unusable

@opencode You know that free things always come with limitations

@opencode I fixed it:

@opencode Fixed what? The post was about API.

@opencode Partially misread the post, my bad. But, We fixed attention and numerical issues in vLLM to get MiMo’s DFlash speculative decoding working correctly in our tests. I wonder if their serve has a similar issue.

@opencode I thought I was the only one.

@opencode thats pain to use

@opencode With mimo 2.6 flash on opencode go (not free), it seems fast enough

@opencode True man quite slow 😔

@opencode Change harness.

@opencode What harness is this

@opencode Around 31 t/s via @CommandCodeAI, probably the API not handling so many people. Seems a pretty solid model to be my Hermes main brain for overall generic tasks so far despite being a bit slow. If @XiaomiMiMo improve speeds, MiMo V2.6 Pro + Opus 5.5 new gold tier for Hermes?

@opencode Wooow That speed belongs only in local with 5+ agents running... Not for an API model

@opencode route to bad region. in china it's has 50-70 tps

@opencode I got about 14 tok/s. On RTX 5090 with 64gb DDR 5 for Flash version Q2

@opencode Ouh.. ;)

@opencode Huh. So you think people is okay? Or you’re still checking if it craps our garbage tokens too?

@opencode qual app é esse?

@opencode what overlay is this?

@opencode Which harness is this

@opencode are you sure you are not hooked up to your own local AI models? :P

@opencode rookie numbers 😎

@opencode What is this harness?

@opencode any non-local llm is just a confounding factor .. which is so hard to take in this space :)

@opencode Use it on Mimo endpoints, faster and cheaper. I bet cache management is more optimized there.

@opencode Is it average or right in the middle?

@opencode @grok which harness is it?

@opencode 20-24 tok/s on average can hide a lot of waiting in opencode when it keeps dipping below 10

@opencode Did you try it on Nous?

@opencode No

@opencode Free will always be slow to support more users.

@opencode What about intelligence, tone and output quality?

@opencode Don’t use opencode go, it’s free for now which meanings many people are using it

@opencode 20 tok/s dipping under 10... flash in the name, buffering in the soul. local speed checks still matter

@opencode How is the limit?