Meta releases Muse Spark Open, 816B-parameter open-weight model
Muse spark is out now and it mogs every other open weight model research.meta.ai/blog/muse-sp...
philpax16GLM-5.3 is smaller and better at SWE. pretty neat w/ the other benchies though
Latte macchiato5This seems comically massive for the results it actually gets, ngl. Especially it losing out to GLM 5.3 in many parts makes it a difficult sell. It's good, but I'm like... whelmed, I suppose. Does look good when compared against Inkling or Mistral, so that's something at least.
Sekoia2Bigger than GLM-5.3 for worse results in the more interesting benchmarks is imo only worth it if it has a much much better arch for tps and/or size reduction
Bitey2They seem to have gone right to an Apple-friendly optimization. Really you can see why Apple stock is so healthy, any small co or dept office could could buy two M5 Ultra clusters for $70K today and have most AI needs met. $70K has been 'dept server' price for decades, now apple can just step in
Fennec Fox Fanfare1at the very least, this is good for commercial users that want open-weights but can't use Chinese models (there's definitely some "we have federal contracts so we can't buy anything that touches GLM or Kimi, it might infect us with the commie cooties" going around in industry)
aly5jesus meta is finally cooking
