Apple is in talks with PrismML, a Khosla Ventures-backed startup, over using its compression technology to run more AI directly on iPhones. PrismML claims it can dramatically shrink model size while preserving performance.
March 31, 2026 · PrismML
A 27-Billion-Parameter Model — Running Directly on an iPhone
Caltech-born, Khosla-backed startup PrismML crushes AI models to a single bit per weight (+1 / −1) with its "1-bit Bonsai" family — and Apple is now in talks to bring the compression tech to on-device Apple Intelligence.
1-bit
Every weight is +1 or −1 — native end-to-end, no high-precision escape hatch
1.15GB
Memory for flagship Bonsai 8B (8.2B params) — 12–14× smaller
$16.25M
Seed round — Khosla Ventures, Cerberus, Google/Caltech compute
Qwen 3.6 compressed to fit a phone
Alibaba's 27B open model shrunk from ~54 GB down to under 4 GB — column height ∝ memory size
→
under 4 GB
1-bit compressed
≈ 13× smaller — cloud-scale intelligence, now local
The Bonsai family footprint
On-device memory by model (GB)
Bonsai 8B · 8.2B params 1.15 GB
Bonsai 4B · 4B params 0.5 GB
Bonsai 1.7B · 1.7B params 0.24 GB
Bonsai 8B throughput
iPhone 17 Pro 40 tok/s
iPhone 17 Pro Max 44 tok/s
M4 Pro Mac 131 tok/s
RTX 4090 368 tok/s
Bonsai 8B benchmarks
Average 70.5
MMLU Redux 65.7
GSM8K 88.0
HumanEval+ 73.8
Energy: 0.068 mWh / token (iPhone 17 Pro Max)
What excites developers
Runs an 8B-class model on an iPhone — even older handsets
Strong size-to-speed balance; privacy & low latency
Apache 2.0 weights; MLX + llama.cpp CUDA support
Bonsai Image 4B: up to 8× compression, 5.6× speedup at ~95% fidelity
What observers caution
Lags full-precision models on complex reasoning tasks
Speed varies significantly by implementation
Full payoff may depend on future dedicated hardware
Apple talks are early — outcome still undecided
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…