PrismML wuxuu soo saaray model yar oo AI ah oo ku habboon mashiinada iyo taleefannada
Shirkadda PrismML ayaa maanta soo saartay model cusub oo ay ku magac baxday Bonsai 2 27B, kaas oo ka dhigay modelka Qwen3.8 27B mid yar oo kaliya 5.9 GB, isagoo ku filan mashiinada iyo taleefannada casriga ah.
Tusaale cusub oo ka dhigaysa AI-ka mid yar oo ku habboon mashiinada
PrismML ayaa soo saartay modelka Bonsai 2 27B, kaas oo ka dhigay modelka Qwen3.8 27B mid yar oo kaliya 5.9 GB, isagoo ka dhigay mid ku habboon mashiinada iyo taleefannada casriga ah. Tani waxay ka dhigaysaa modelka mid yar oo 9x ilaa 10x ka yar kan hore.
Shirkadda PrismML ayaa sheegtay in modelkan cusub uu ka dhigay modelka Qwen3.8 27B mid yar oo kaliya 5.9 GB, isagoo ka dhigay mid ku habboon mashiinada iyo taleefannada casriga ah. Tani waxay ka dhigaysaa modelka mid yar oo 9x ilaa 10x ka yar kan hore.
Tignoolajiyadda PrismML iyo faa'iidooyinkeeda
PrismML waxay isticmaashaa tignoolajiyad loo yaqaan 'ternary' weights, taas oo ka dhigaysa modelka mid yar oo kaliya 5.9 GB, isagoo ka dhigay mid ku habboon mashiinada iyo taleefannada casriga ah. Tani waxay ka dhigaysaa modelka mid yar oo 9x ilaa 10x ka yar kan hore.
Shirkadda PrismML waxay isticmaashaa tignoolajiyad loo yaqaan 'ternary' weights, taas oo ka dhigaysa modelka mid yar oo kaliya 5.9 GB, isagoo ka dhigay mid ku habboon mashiinada iyo taleefannada casriga ah. Tani waxay ka dhigaysaa modelka mid yar oo 9x ilaa 10x ka yar kan hore.
Hadda iyo horumarka soo socda
PrismML waxay sheegtay in modelkan cusub uu ka dhigay modelka Qwen3.8 27B mid yar oo kaliya 5.9 GB, isagoo ka dhigay mid ku habboon mashiinada iyo taleefannada casriga ah. Tani waxay ka dhigaysaa modelka mid yar oo 9x ilaa 10x ka yar kan hore.
Shirkadda PrismML waxay isticmaashaa tignoolajiyad loo yaqaan 'ternary' weights, taas oo ka dhigaysa modelka mid yar oo kaliya 5.9 GB, isagoo ka dhigay mid ku habboon mashiinada iyo taleefannada casriga ah. Tani waxay ka dhigaysaa modelka mid yar oo 9x ilaa 10x ka yar kan hore.
PrismML, an AI lab founded by Caltech researchers, has released Bonsai 2 27B, a compressed version of Alibaba's Qwen3.8 27B model that fits in just 5.9 GB while matching 98% of the original's benchmark scores. The startup is betting that powerful reasoning models do not need to be large to be useful.
A Tiny Model With Big Claims
PrismML, led by Caltech professor Babak Hassibi, has raised a $22.25 million seed round and is backed by Khosla Ventures, Cerberus Capital, and Caltech itself. The company is making reasoning models small enough to run on personal computers and smartphones, a move that could reshape how users access artificial intelligence. Ion Stoica, co-founder of Databricks and director of Berkeley's Sky Computing Lab, serves as an advisor to the startup.
The new release, Bonsai 2 27B, compresses the widely used Qwen3.8 27B model down to just 5.9 GB, representing a 9x to 10x reduction in memory compared with the original. This size is small enough to fit on a PC and possibly a high-end smartphone, making advanced AI capabilities accessible to a much wider audience without relying on cloud infrastructure.
Near-Perfect Performance Through Ternary Compression
Bonsai 2 27B matches 98% of Qwen's aggregate benchmark scores, an improvement from the first Bonsai model released in March, which achieved 95%. The original Bonsai has been downloaded over 11 million times, while PrismML's even smaller models have attracted another 2.6 million downloads, demonstrating strong market interest in the technology.
The compression technique relies on 'ternary' weights, which simplify each weight from 16 bits to just three values: +1, −1, or 0. CEO Babak Hassibi noted that while perfect benchmark parity may remain elusive, the 2% degradation is unlikely to meaningfully affect real-world performance. He added that as model size grows, there is more room to compress without losing intelligence.
Looking Ahead to Larger Models and On-Device AI
PrismML's next goal is to apply its compression technique to models in the several-hundred-billion-parameter range, which Hassibi expects will be easier to compress while retaining intelligence. He stated that larger models offer more room for compression without losing performance, suggesting that achieving 100% benchmark parity may become more feasible as the technology matures.
Stoica expressed excitement about the potential for advanced models to run directly on users' devices, emphasizing that such AI would be free and private since it would not require sending data to the cloud. Hassibi declined to comment on rumors of talks with Apple, but the company's vision of bringing capable AI to personal devices remains a central pillar of its strategy.
Ilaha iyo xuquuqda sawirka
Sawir: TechCrunch Xigasho
