Fitting a 27B Model on an iPhone
ElmBark · reddit · 2026-07-17
PrismML managed to create an iPhone-runnable version of Bonsai (based on Qwen3.6-27B) by compressing its weights down to 1-bit.
- The original model was roughly 54GB, but the quantized version is only 3.9GB, easily fitting on a phone.
- It utilizes true binary quantization: each weight has just 1 sign bit, sharing a single FP16 scale per 128 weights.
- Unusually, even the embeddings, attention/MLP projections, and LM head were binarized.
- Across 15 benchmarks, the average score dropped from 85.1 (FP16) to 76.1, retaining about 89.5% of the performance.
- Math tasks saw the best retention (91.7%), while knowledge and reasoning experienced more noticeable degradation.
- Memory footprint remains low: 5.2GB for a 4K context and 6.8GB for a 100K context (using a 4-bit KV cache).
The original post also notes that this version can already run on an iPhone 15 Pro Max via Atomic Chat.
More from Infra
- Resin lets an agent build standalone LLM inference engines; vLLM and SGLang optimize DeepSeek kernels differently — AccBalanced · 2026-10-27
- Local benchmark: Flint matches Qwen3.6 quality at 50/52 while running 3.6x faster on M4 Pro — RichardSocher · 2026-10-27
- Alchemy adds effect/Workflow support across Durable Objects, Workflows and Lambda Durable Functions — samgoodwin89 · 2026-10-27
- Alchemy's Lambda Web Functions run plain Node web servers on Lambda, no OpenNext needed — samgoodwin89 · 2026-10-27
- Alchemy Cuts Cloudflare Bindings CPU Time by Up to 166× Via Profiling — samgoodwin89 · 2026-10-27
- Foldkit Adds SSR Support With a Few Lines of Config on Cloudflare and AWS — samgoodwin89 · 2026-10-27