The 1-bit llm is here and it's not a paper toy. prismml's bonsai, out september 17 with $16.25m…
the 1-bit llm is here and it's not a paper toy. prismml's bonsai, out september 17 with $16.25m behind it, runs an 8b model in 1gb of ram against 16gb for fp16 and hits 44 tokens a second on an iphone.
the gap between cloud and pocket just closed.
Context
PrismML announced Bonsai 8B on March 31, 2026, and HPCwire reported on April 6, 2026 that it emerged from stealth with a $16.25 million seed round. The model card says 1.28 GB of parameter memory against 16.38 GB at FP16, 12.8x smaller, and 44 tokens per second on an iPhone 17 Pro Max through MLX Swift in PrismML's own fork, with upstream support pending. Its table shows 14 tokens per second for a 4-bit baseline on that phone. It lists end-to-end 1-bit weights across embeddings, attention, MLP projections and the LM head, in a g128 format where every 128 weights share one FP16 scale, on a Qwen3-8B dense architecture.
The note says out September 17. The first-party date for Bonsai 8B is March 31, 2026, and September 17 is the date of a different model, Bonsai 2 27B. The funding is a seed round reported in April. The 1 GB figure rounds 1.28 GB of parameter memory only, and does not include the KV cache or runtime memory, so it is not total operation in 1 GB of RAM. The 16 GB baseline is FP16, which does not fit on the phone, and the 44 tokens per second comparison is against a 4-bit baseline, about 3.1x. Activation precision was not stated in the text read. The 70.5 average benchmark score and the Caltech research claim are vendor-reported, and the paper was not read. All figures are the vendor's and not independently audited.
Related work
- PrismML shipped Bonsai 2 27B on September 17 ↗The earlier note on the later ternary 27B model, which is the September 17 release.
- Ternary Bonsai 2 is a 27B model ↗A second earlier note on the same later model.
Watch next
- Independent replication of the 44 tokens per second, the Caltech paper, and whether upstream MLX merges the 1-bit kernels.
Sources
- Hugging Face: prism-ml/Bonsai-8B-mlx-1bithuggingface.co
- PrismML: launches 1-bit AI model (March 31, 2026)prismml.com
- HPCwire: PrismML emerges from stealth (April 6, 2026)hpcwire.com
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 21 September 2026 at 21:35 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 →