Qwen3.8 27B is an open model from Alibaba, released August 15, that runs locally and scores near the top of third-party benchmark tables. PrismML has since used it as the base for Ternary Bonsai 2 27B, a compressed 5.9GB build that keeps most of its benchmark scores.
Alibaba released Qwen3.8 27B on August 15 as an open model that runs locally. Two days later, third-party analysis from Artificial Analysis put it at an Intelligence Index of 52, tying GPT-5.6 Luna, and an Agentic Index of 51, above GPT-5.6 Terra and Claude Opus 4.8. Those numbers made it the rare open model sitting level with closed frontier releases, and with no API fees attached.
A month later, PrismML turned that base into something smaller. Ternary Bonsai 2 27B ships at 5.9GB, one-ninth the size of the model it is built on, and holds 98.2% of Qwen3.8 27B's benchmark scores. PrismML lists gains in reasoning, coding, vision, and long-horizon agent performance over the earlier Bonsai 27B. The compression matters because it puts local execution within reach on 8GB of VRAM.
The licensing and distribution are settled. PrismML released Ternary Bonsai 2 27B under Apache 2.0 and published it on Hugging Face, and the model runs on Mac, iPhone, and iPad. The base model and its compressed derivative now occupy two positions in the same local stack: the full-weight open release and a small-footprint build aimed at hardware that cannot hold the original.
Synthesized by Yomimono from the cited Yomimono stories below, each itself
sourced, then editorially reviewed. Every
fact links the story it came from.
Sep 18
PrismML announced Ternary Bonsai 2 27B, a compressed model built on Alibaba's Qwen3.8 27B. It ships at 5.9GB, one-ninth the size of the base model, and retains 98.2% of Qwen3.8 27B's benchmark scores. PrismML lists gains in reasoning, coding, vision, and long-horizon agent performance over the earlier Bonsai 27B. The model is released under the Apache 2.0 license and published on Hugging Face. PrismML says it runs on Mac, iPhone, and iPad.
Aug 18
Alibaba released the open model Qwen3.8 27B on August 15. Third-party analysis from Artificial Analysis on August 17 gives it an Intelligence Index of 52, tying GPT-5.6 Luna, and an Agentic Index of 51, surpassing GPT-5.6 Terra and Claude Opus 4.8. The model runs locally.