This Week in Open Models: Bonsai 27B Runs on an iPhone (3.9 GB) and Nemotron 3 Embed Takes the #1 RTEB Spot

TL;DR: PrismML's Bonsai 27B squeezes frontier-grade 27B reasoning into 3.9 GB via 1-bit quantization; NVIDIA's Nemotron 3 Embed 8B just took #1 on the RTEB retrieval benchmark — both free under Apache 2.0.

Bonsai 27B — a 27B model that fits on an iPhone. PrismML released Bonsai 27B on July 14: a 1-bit quantized version of Qwen3.6 27B that runs at 11 tok/s on an iPhone 17 Pro at just 3.9 GB. A 1.58-bit ternary variant at 5.9 GB targets laptops. The model handles multi-step reasoning, structured tool use, vision, and 128K-token context. Apple is reportedly in talks with PrismML about the compression technology. → prismml.com

NVIDIA Nemotron 3 Embed — new #1 embedding model. Released July 17, Nemotron 3 Embed is an open, commercially-licensed collection of three models. The flagship 8B checkpoint (Nemotron-3-Embed-8B-BF16) scores 78.5% on RTEB — #1 overall, across 34 languages, with a 32,768-token maximum sequence length. Two efficient 1B variants (BF16 and Blackwell-optimized NVFP4) target production RAG at lower cost. Weights on HuggingFace now. → HuggingFace blog

Why it matters: 1-bit compression is making 27B-class models viable on commodity hardware, and the RTEB leaderboard now has an open model at the top — both trends reduce lock-in to closed APIs for inference and retrieval.

Sources: Bonsai 27B release — PrismML · Nemotron 3 Embed #1 on RTEB — HuggingFace Blog · Bonsai 27B on 9to5Mac