SHELFCRITTER AI Digest

COMPANY

PrismML

prismmledge-ai1-bit-llmtopic-note

PrismML

AI startup founded by Caltech researchers, focused on extreme model compression for edge deployment. Emerged from stealth in April 2026 with a $16.25 million seed round and the Bonsai 1-bit LLM family — models where every weight is represented by its sign ({-1, +1}) plus a shared scale factor, trained natively at 1-bit precision end-to-end. The flagship Bonsai 8B fits in 1.15 GB and runs 8x faster than full-precision equivalents on edge hardware. Models released under Apache 2.0.

Timeline

  • 2026-04-06-AI-Digest — PrismML emerged from stealth with $16.25M seed round and the open-source Bonsai 1-bit LLM family (8B, 4B, 1.7B) under Apache 2.0, enabling competitive edge inference at a fraction of typical memory requirements.
  • 2026-07-15-AI-Digest — PrismML releases Bonsai 27B — 1-bit and ternary quantisations of Qwen3.6-27B that run on-device (HN: 501 pts / 186 cmts). Ternary retains ~95% of FP16 quality across 15 benchmarks; 1-bit retains ~90%. Distinct from the natively-1-bit Bonsai family (April 2026): today’s release applies the compression pipeline to an external open-weight base model rather than training a 1-bit model from scratch. The “27B on a phone” framing needs the quantisation caveat — this is a compression feat, not a new pretrain — but on-device 27B-class inference materially expands what offline LLM assistants can promise.

Context

Timeline

  • 2026-07-16-AI-Digest — PrismML ships Bonsai 27B as a fully open reasoning model — a ternary / 1-bit-quantized derivative of Qwen3.6-27B — running on-device on an iPhone. Narrow read: the substantive note is that Bonsai 27B is derivative — a compression story of a Chinese open base rather than an independent open reasoning model, so treat it as more evidence for the Chinese-open-frontier leg of the three-way US/Chinese open/US closed split, not for the US-open below-frontier leg. Structural read: paired with DeepMind‘s verification-bottleneck framing the same digest, PrismML is compounding on the on-device compression axis while frontier labs converge on “the bottleneck is downstream of generation” — two independent AI-for-science-adjacent framings landing the same news cycle.

  • 2026-09-18-AI-Digest — PrismML ships Bonsai 2 27B — a ternary-quantised ({−1, 0, +1}) derivative of Qwen 3.8 27B that fits in 5.9 GB, carries a 262K context, ships under Apache 2.0, and retains ~98.2% of the Qwen 3.8 27B baseline on an aggregate 20-benchmark score of 83.9. The company also disclosed a $22.25M seed (Khosla, Cerberus, Google, with Samsung backing) not previously surfaced in the corpus. Load-bearing framing: the baseline being matched is the same-size Qwen 3.8 27B, compressed to a phone-deployable footprint at ~98% quality retention, not a headline-friendly cross-size collapse; earlier reporting conflating Bonsai 2 with a “27B compresses to a 243B baseline” claim is wrong. Distinct release from Bonsai 27B‘s July 2026 quantisation of Qwen3.6-27B — Bonsai 2 targets the newer Qwen 3.8 base and pushes to 5.9 GB. Benchmarks are PrismML-reported; treat as vendor-claim until an independent bench (Aider, LM Arena) posts numbers. HN traction 326 pts / 106 cmts.

Key Developments

  1. Bonsai 2 27B ships as ternary-quantised Qwen 3.8 27B derivative (September 17, 2026): 5.9 GB / 262K context / Apache 2.0, aggregate ~98.2% of Qwen 3.8 27B baseline (20-benchmark suite, 83.9). Distinct from Bonsai 27B’s July 2026 Qwen3.6-27B quantisation — Bonsai 2 targets the newer base and pushes to a smaller ternary footprint. PrismML $22.25M seed disclosed same news cycle (Khosla, Cerberus, Google, Samsung).