TechCrunch on 17 September described PrismML shipping Bonsai 2 27B as a squeezed take on Alibaba's Qwen3.8 27B in a 5.9 gigabyte pack. The company claims a 9 times to 10 times memory cut and 98 percent of the Qwen aggregate on its own sheet. Hassibi works on compression at Caltech; Stoica advises. The first Bonsai, PrismML said, passed 11 million downloads. No outside tally of the 98 percent line is in the report.
TechCrunch, on 17 September, described PrismML shipping Bonsai 2 27B as a squeezed take on Alibaba's Qwen3.8 27B. The pack size given is 5.9 gigabytes, small enough, the company says, for a desktop and maybe a high-end handset. Memory is claimed down by 9 times to 10 times. Readers should treat every percentage here as Anthropic or PrismML speaking in their own post, not as an audited outside ledger.[1]
PrismML put the compressed model at 98 percent of the Qwen aggregate on its own sheet. TechCrunch wrote the boast down and left the absent outside yardstick visible. Readers should treat every percentage here as Anthropic or PrismML speaking in their own post, not as an audited outside ledger.[1]
Hassibi, the founder, works on compression at Caltech. Stoica sits as an adviser. The first Bonsai pass, TechCrunch wrote, crossed 11 million; smaller models crossed 2.6 million. Hassibi put the next band at a few hundred billion parameters. The retained page also cites a 5.9-gigabyte footprint for that first pass. This slot has no second publisher.[1]
