Releases·Model Release
PrismML brings its tiny LLMs to Qualcomm-powered smart glasses
At Qualcomm's Snapdragon Summit, the chipmaker demonstrated PrismML's 1-bit Bonsai LLM running locally on smart glasses built on the Snapdragon AR1 Gen 1 Platform. The 2-billion-parameter model, tuned for vision and language, lets wearers ask about what they are seeing in real time, advancing PrismML's goal of open-weight on-device AI.
0 engagement1 sourceThu, Sep 24, 2026, 07:00 PM
AI Overview
- PrismML, founded by Caltech researchers and advised by Ion Stoica, showed its 1-bit 2B-parameter Bonsai LLM running locally on Snapdragon AR1 Gen 1 smart glasses at Snapdragon Summit
- The glasses-tuned vision-language model lets wearers ask about what they're seeing in real time; PrismML shrinks models 4x with near-full benchmark performance
- No smart glasses running PrismML models have been announced yet
Related
Releases·Model Releasefeatured
PrismML releases Bonsai 27B, a 1-bit quantized 27B-parameter LLM that runs on a phone
PrismML released Bonsai 27B, a 1-bit quantized version of Qwen3.6 27B that shrinks from 54GB to 3.8GB while retaining 90% of benchmark performance. It is the first 27B-class model capable of running locally in a browser or on a smartphone via custom WebGPU kernels, enabling on-device inference without cloud dependency.
PrismML released Bonsai 27B, a 1-bit quantized version of Qwen3.6 27B that shrinks from 54GB to 3.8GB while retaining 90% of benchmark performance. It is the first 27B-class model capable of running locally in a browser or on a smartphone via custom WebGPU kernels, enabling on-device inference without cloud dependency.
Updated Fri, Jul 24, 2026, 07:02 PM — Bonsai 27B now runs on iPhone (3.9GB) and in browser via WebGPU; Ollama desktop app released July 24.
5413.2k engagement10 sourceshackernews, reddit, rss +1Tue, Jul 14, 2026, 05:50 PM
News·Company News
Apple in talks with PrismML to shrink AI models for on-device iPhone inference
Apple is reportedly in discussions with startup PrismML, which specializes in compressing AI models to run efficiently on mobile devices like the iPhone. The talks suggest Apple is seeking to bring more advanced on-device AI capabilities without relying on cloud processing.
Apple is reportedly in discussions with startup PrismML, which specializes in compressing AI models to run efficiently on mobile devices like the iPhone. The talks suggest Apple is seeking to bring more advanced on-device AI capabilities without relying on cloud processing.
374 engagement2 sourcesredditWed, Jul 15, 2026, 12:23 PM
Social·Community
Ternary Bonsai 2 27B (1.75bpw) vs. Gemma 26B-A4B MoE
A developer benchmarked prism-ml's Ternary-Bonsai-2-27B quantized model (1.75 bpw, 5.95 GB) against Gemma 4 26B-A4B on an 8 GB Turing GPU for dialogue speaker attribution in an audiobook pipeline. Despite fitting a 27B-class model fully in VRAM, Bonsai scored 93.3% attribution accuracy but ran 5-10x slower and showed systematic narrator and intonation errors; the tester declined to adopt it.
A developer benchmarked prism-ml's Ternary-Bonsai-2-27B quantized model (1.75 bpw, 5.95 GB) against Gemma 4 26B-A4B on an 8 GB Turing GPU for dialogue speaker attribution in an audiobook pipeline. Despite fitting a 27B-class model fully in VRAM, Bonsai scored 93.3% attribution accuracy but ran 5-10x slower and showed systematic narrator and intonation errors; the tester declined to adopt it.
25 engagement1 sourceredditSat, Sep 19, 2026, 01:54 AM
News·Company News
Google AI Edge team shrinks Gemma to run on Raspberry Pi, highlights worsening RAM constraints
Cormac Brick of Google AI Edge argues that RAM, not compute, is the primary constraint for edge AI, noting that phone makers are shipping less RAM and a 6GB Raspberry Pi now costs 2.5x its launch price. His team has quantized a 2B-parameter Gemma model to 2.9 bits per weight, achieving ~8 tokens/second on a Raspberry Pi and fast enough vision processing on Qualcomm NPUs. They also target tiny models from 500M down to 50M parameters for older laptops and cheap devices.
Cormac Brick of Google AI Edge argues that RAM, not compute, is the primary constraint for edge AI, noting that phone makers are shipping less RAM and a 6GB Raspberry Pi now costs 2.5x its launch price. His team has quantized a 2B-parameter Gemma model to 2.9 bits per weight, achieving ~8 tokens/second on a Raspberry Pi and fast enough vision processing on Qualcomm NPUs. They also target tiny models from 500M down to 50M parameters for older laptops and cheap devices.
113 engagement1 sourceyoutubeSat, Jul 25, 2026, 05:00 PM
Social·Community
Users report Bonsai 27B model fails at simple coding tasks
A Reddit user reports that the Bonsai 27B model, a ternary-precision model from PrismML, is "pretty useless" for simple coding tasks and story generation. The model cuts off its thinking process and fails to respond with results, even with increased context size.
A Reddit user reports that the Bonsai 27B model, a ternary-precision model from PrismML, is "pretty useless" for simple coding tasks and story generation. The model cuts off its thinking process and fails to respond with results, even with increased context size.
32 engagement1 sourceredditWed, Jul 15, 2026, 05:32 AM