The reasoning model shrank down to a phone. prismml's ternary bonsai 2 27b, released september 17,…
the reasoning model shrank down to a phone. prismml's ternary bonsai 2 27b, released september 17, compresses qwen3.8 into 5.9 gigabytes while keeping 98% of benchmark performance, with 13 million downloads across the lineup.
the gpu is now optional.
Context
PrismML's post of 17 September 2026 introduces Bonsai 2 27B, based on Qwen3.8 27B with ternary weights at 1.76 effective bits per weight and a total footprint of 5.9GB, more than 9x smaller, with 98.2 percent of aggregate performance (83.9 aggregate), 262K context, up to 143 tokens per second on an NVIDIA RTX 5090 and 46.8 on an M5 Max, weights under Apache 2.0, running on NVIDIA GPUs via CUDA and on Apple devices (Mac, iPhone, iPad) via MLX. The Hugging Face card lists FP16 at about 54 GB, ideal ternary 5.8 GB, a GGUF pack at 5.95 GB and MLX 2-bit at 7.67 GB.
The 5.9GB and 98.2 percent are first-party and vendor-reported, not independent. 5.9GB is the headline footprint and not one file for every runtime, since the Apple MLX 2-bit file is 7.67 GB. The post names iPhone as a supported device family but no phone memory requirement, speed or demo was found; the measured devices are an RTX 5090 and an M5 Max laptop. The 13 million downloads was not found: a third-party page shows 3.1 million for one earlier repo and no sum across repos. The GPU is now optional is the author's take, since the numbers are on a GPU and a laptop chip.
Watch next
- PrismML's per-benchmark table, any iPhone-specific numbers and download counts per repo.
Sources
- Bonsai 2 27B (PrismML, 17 Sep 2026)prismml.com
- Ternary-Bonsai-2-27B MLX 2-bit (Hugging Face)huggingface.co
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 21 September 2026 at 12:41 IST. Sources are the papers and datasets the note draws on.
View the original post ↗Embed this note
More notes
The air is now being asked to keep its own ledger
the air is now being asked to keep its own ledger: ecmwf’s aifs compo becomes the first ai model to forecast atmospheric composition globally every three hours, cleanair simulates 365 days of pm2.5 over china in ten seconds, and a unified framework maps six pollutants at one kilometer across the whole country. the air now files its own composition report.
read the note →The current is now being asked to draw its own map
the current is now being asked to draw its own map: china’s langya 2.0 predicts six ocean phenomena including internal waves and mesoscale eddies, a deep net called wenhai resolves eddies globally with air sea flux formulas built in, and scripps infers surface currents from the way temperature patterns deform in satellite images. the ocean now files its own circulation report.
read the note →The soil is now being asked to report its own carbon
the soil is now being asked to report its own carbon: a nix color sensor paired with generative data augmentation predicts soil organic carbon without a lab, random forest drives 74 percent of soil health mapping studies, and sentinel 2 tracks five year carbon change across france and italy from 922 samples. the dirt now files its own carbon account.
read the note →