LecturaInvestigaciónRadarMarco de inversión
Iniciar sesión / Registrarse
← Volver a la lectura
NoticiasTema · 模型压缩与效率Material2 min de lecturaEditorial de DisconfirmAI

AI Startup PrismML Releases Bonsai 2: Compresses 27B Model to 5.9GB with 98% Self-Tested Performance Retention

PrismML launched its second-generation compressed model, Bonsai 2, which uses ternary weight quantization to shrink the Qwen3.8 27B model to 5.9GB for local execution on PCs and phones, claiming it retains 98% of the original's benchmark score.

2026-10-20 09:33 GMT+8

PrismML, an AI startup founded by researchers from Caltech, released Bonsai 2 27B on October 20, 2026. Based on Alibaba's open-source Qwen3.8 27B model, this second-generation compressed model utilizes a unique 'ternary' weight compression technique to reduce the total capacity to 5.9GB, enabling direct execution on personal computers or high-end smartphones.

According to Babak Hassibi, founder and CEO of PrismML, the company's approach aims to preserve performance almost entirely. Internal benchmark results published by the company indicate that while memory usage was reduced by 9 to 10 times compared to the original version, Bonsai 2 27B achieved 98% of the composite benchmark score of the original Qwen3.8 27B, an improvement over the first-generation Bonsai model's 95% score reported six months prior.

Technically, the team simplified each 'weight'—which typically requires 16 bits—to one of three values: +1, -1, or 0, significantly reducing storage requirements. Hassibi acknowledged that compression inevitably impacts performance to some degree, and whether 100% parity can be achieved remains to be seen. He also noted that benchmark scores do not fully reflect real-world task performance, as final accuracy is influenced by software environments and frameworks.

PrismML plans to apply this compression technology to models with hundreds of billions of parameters in the coming months. For consumers and small-to-medium enterprises, high-performance small-capacity models offer free, zero-overhead access to AI capabilities anywhere, anytime, while enhancing privacy by eliminating the need to transmit data to the cloud.

Iniciar sesión / Registrarse
LecturaInvestigaciónRadarMarco de inversión

Fuentes

  • techorange.com↗
Tema

模型压缩与效率

AI行业正通过架构创新与极致压缩挑战传统大模型的算力瓶颈。美国初创公司Inception Labs于2026年10月20日发布旗舰扩散式大语言模型Mercury 3,宣称其生成速度达每秒1000个token,推理成本显著低于Anthropi…

2 pasos2 piezasActualizado 2026-10-21 04:52

Cómo llegó hasta aquí

  1. 2026-10-20Inception Labs发布扩散式大语言模型Mercury 3,CEO Stefano Ermon称其速度达1000 tokens/秒,价格比竞品低35%,并声称在代理基准测试中表现优于Haiku 5.5等模型。thedeepview.com ↗
  2. 2026-10-20PrismML发布Bonsai 2 27B,基于Qwen3.8 27B进行三元权重压缩,体积降至5.9GB,基准测试得分达原版98%,旨在支持本地设备运行。techorange.com ↗

Lo que aún está pendiente

  • Mercury 3的扩散架构在处理长上下文逻辑推理时,是否能持续保持与自回归模型相当的质量?
  • Bonsai 2等压缩模型在实际用户场景中的体验损失(如幻觉率增加)是否真的如基准测试所示微乎其微?
  • 随着参数规模扩大,压缩模型能否真正实现100%性能保留,还是存在不可逾越的信息熵极限?

También en este hilo

  • AI startup Inception launches Mercury 3, a diffusion-based LLM claiming ~1,000 tokens per second generation speed2026-10-20
Leer el tema completo →
Continuar investigando →