Developers reported 44.3 tokens per second on code from a 105 GB MiMo pack running on one machine with 128 GB of memory. The pack fit, and it ran at a reported speed. But there are no task scores for that exact compressed pack, so the speed result leaves a basic question open: how well did it answer?
A 105 GB Pack Ran on One Machine
Developers tested compressed MiMo V2.6 Flash and GLM 5.3 Flash packs on one Ryzen AI Max Plus 395 with 128 GB of memory. The MiMo pack was 105 GB; the GLM pack was 99.7 GB. Both fit on that machine in the developers’ tests.
GLM’s size needs a little context. ZI lists it at 320 billion parameters, with only 18 billion active for each token. But all its experts still have to stay in memory. The active count describes how much of the model works on a token; it does not make the rest of the compressed pack disappear from memory.
For MiMo, the developers reported 28.9 tokens per second without speculation. On code, it reached 44.3 tokens per second with speculation: a small draft model guessed ahead, and the larger model checked several tokens in one pass. They reported 26 to 30 tokens per second for GLM with speculation.
Those are useful fit and speed observations, with a clear limit: the speeds have not been independently reproduced. They also describe different runs. MiMo’s 44.3 figure came from code with speculation, while its 28.9 figure came from a run without speculation.
Speed Leaves the Task Question Open
The quality cards measure how often each compressed pack picks the same next token as the official weights. That comparison can tell a reader something about next-token agreement. It gives no task scores for these exact packs.
That distinction matters because the reported numbers answer different questions. The pack sizes tell us what fit on the 128 GB machine. The speed runs tell us how quickly the tested setups produced tokens. Neither result tells us what the compression preserved on tasks.
There is another limit on the MiMo speed result: its tests used temperature zero, which MiMo’s own card advises against. The reported speed is still the reported speed, but the card’s advice is part of how to read that test. For now, the compressed MiMo and GLM packs have reported fit and speed results, while their task performance remains unknown.
The 64 GB Spark Has a Different Fit Limit
Nvidia’s coming 64 GB DGX Spark costs $4,999. It has the same GB10 chip as the 128 GB version, but less memory, and Nvidia puts its model ceiling at up to 100 billion parameters. The speaker says heavily compressed GLM 5.3 Flash and Deepseek V4 Flash builds still will not fit.
That memory difference changes how to read other Spark results. Artificial Analysis tested eight agent tasks on the 128 GB Spark. Those results came from the larger-memory machine, so they do not establish what fits or runs on the coming 64 GB version.
A separate hands-on video reported the RTX 5090 running more than 3.5 times as fast as the other tested machines. That comparison covered models that fit in its 32 GB. It cannot settle what happens with a compressed pack too large to fit there.
The reported MiMo and GLM runs show that large compressed packs can fit and run on the tested 128 GB machine. The other hardware results have their own memory limits. And for the exact MiMo and GLM packs with reported speeds, task scores remain unreported.