Meta recently released Muse-Glimmer, an open-weights/open source model sized for commodity hardware.
I downloaded the Ollama muse-glimmer:30b-mxfp8-dflash
edition onto my 64GB M2 Ultra Mac Studio. mxfp8
is a less-aggressive compression model than I usually employ on that machine (nvfp4
). dflash
uses a predictive algorithm to help improve speed, supposedly.
I then had Muse-Glimmer run the same documentation-review prompt that I had Qwen 3.6 run previously, covered in my preceding issue. Muse-Glimmer ran more than twice as long and came up with a small fraction of what Qwen did.
Muse-Glimmer may have uses, but I suspect that few of those will tie into coding agents.
I might give NVIDIA's Nemotron 3.5 Lightning a shot, though NVIDIA's own benchmarks have it lower at coding tasks than Qwen 3.6.