PrismML
Frontier-class open-weight models compressed to run on a phone
💸 No earnings reported yet
What it is
PrismML builds ultra-dense quantized language models under the Bonsai line. Bonsai 27B compresses Qwen3.6-27B to 1.125 bits per weight — 3.9GB for the 1-bit build, 5.9GB for the ternary — retaining 89.5% and 94.6% of the FP16 baseline respectively, with a 262K context and a 4-bit vision tower for multimodal input. The 1-bit build runs at roughly 11 tokens/sec on an iPhone 17 Pro Max and 66 tok/s on an M5 Max. Smaller Bonsai 8B, 4B and 1.7B models plus Bonsai Image are also available. Weights ship on HuggingFace under Apache 2.0.
How AI plugs in
Alternatives & related tools
Hugging Face
The home of open-source AI
Ollama
Run open LLMs locally
Evo 2
Genomic foundation model for DNA, RNA, and proteins
LG AI Research EXAONE
Korea's frontier open-weight model family.
Llama
Meta's open-weight LLM family, the Western ecosystem anchor.
ModelScope
Alibaba's open model hub and serving platform.
★ Reviews
No reviews yet — be the first.Your rating