Looks like MiMo-V2.6-Flash weights is officially out and is natively QAT!
Very interesting, perfect size for 2x Spark. I’m looking forward to giving it a go.
I followed its training, that was mildly interesting to watch.
Yes and apparently it’s on par with glm 5.3 flash, while smaller and less actives. Only problem, v2. 5 flash wasn’t fast, arch proven to be complex to properly implement and loops were quite present.
The blog has some easier to read benchmarks: MiMo-V2.6 | Xiaomi
Am hoping this runs well on Sparks. Looks interesting.
Probably we will see some early hack from Tony and Mr. Claude tomorrow :)
@tonyd615 batsignal
It says to just follow the Mimo 2.5 recipe. Same architecture.
Worked almost right out of the box with vllm 0.29.0 for me. It is the same family as 2.5.
Has anyone tested this against GLM-5.3-Flash NVFP4/EXL3 on a TP=2 cluster?
Yes, empirically it’s a big step up from GLM-5.3-Flash NVFP4, although to be fair I didn’t directly compare the two.
I’ve been running DeepSeek V4 Flash Vision Exp for the past few weeks, which performed reliably better than the RedHat GLM-5.3-Flash NVFP4 model I tried for coding tasks. MiMo 2.6 Flash is much better again - very similar to Opus 5 performance.
The only downgrade vs DeepSeek v4 Flash Vision Exp is I have to run MiMo 2.6 Flash at 256k context vs 1M. This isn’t a big deal for me.
Why so? No ram, massive kv cache req?
DeepSeek uses an MLA FP4 KV cache, MiMo uses FP8.
Glm 5.3 is bigger, especially with dflash, uses fp8 and we have 1M ctx with 11GB cache pin
This says fp8 will get 1.8M kv cache for tp2. Haven’t tried yet tho.
Sorry I forgot to post guys lol. Been testing it all night but got it standing doing some numbers now. Faster than GLM 5.3 Flash but not sure if its BETTER but def probably willl be better than DS4.1 and Qwen 3.8 MiMo-V2.6-Flash-RL, about 310B total / 12B active MoE DFLASH Running on 2X DGX Sparks PEAK 80 tok/s in CODE
Better than 4.1 - easy goal hahaha. I love 4.1 but hate to give it any work - you send it to buy milk, it will drive around continents to try goat milk, camel milk etc but to see what’s up. Waste of time massively. But Qwen - nice. Qwen punches super hard for 125B model, in 300-400B league
Yea I went back to GLM yesterday it talks to much and over explains too much “Hey just so you know” reminds me of Chatgpt 5.3 and 5.4 on openclaw
GLM on High is just fantastic. Concise and smart. To the point. Let’s see how MiMo stacks against it.