Bonsai 27B — The First 27-Billion Parameter AI Model That Runs on a Phone
PrismML launched Bonsai 27B — a 27-billion parameter AI model that fits in just 5.9 GB and runs on a phone or laptop while retaining high-quality reasoning, tool use, and agentic capabilities.
A Breakthrough in Intelligence Density
PrismML has announced Bonsai 27B, a 27-billion parameter AI model that can run directly on a phone. This model uses ternary weights with FP16 group-wise scaling, allowing it to maintain high-quality reasoning while occupying only 5.9 GB of storage.
The 1-bit version (~3.5 GB) runs directly on an iPhone 17 Pro Max. This means a model with multi-step reasoning, structured tool calls, vision tasks, and computer-use agentic capabilities can work on devices people already own.
Benchmark Performance
Bonsai 27B achieves 80.5% overall on 15 benchmarks compared to 85% for the full-precision Qwen 3.6 27B baseline. Math and coding are nearly untouched — math benchmarks score 93.4% vs 95.3%, and coding at 86.0% vs 88.7%. Tool calling remains within a few points of full precision at 74.0% vs 80.0%.
These results are remarkable because the model occupies 10x less memory than its full-precision counterpart while retaining 95% of its intelligence.
Intelligence Density: A New Metric
PrismML introduces "intelligence density" as a key metric — measuring intelligence per gigabyte. Bonsai 27B delivers 0.53 intelligence per GB, which is 10x more than the full-precision baseline and roughly 2.7x the best available low-bit alternative.
Why This Matters for Agentic AI
The most valuable AI workloads are shifting from single responses to sustained work — assistants that operate real tools, workflows that run unattended, and research that synthesizes dozens of documents. An agent doesn't make one model call, it makes hundreds.
Local execution changes the equation. When the model fits on the device, the marginal cost of a hundred-step loop is zero, and the user's data never leaves the machine. This enables persistent on-device agents, assistants that work offline, and reasoning over private local data.
Hybrid Deployment Architecture
Bonsai 27B enables a new system architecture: hybrid deployments that route non-frontier tasks to the capable local model and reserve frontier cloud models for the hardest steps. This collapses the cost-per-task of agentic systems dramatically.
Implications for Georgian Developers
For Georgian developers and businesses, Bonsai 27B is significant because it democratizes access to advanced AI. No expensive cloud GPUs needed. A capable model that fits on a laptop or phone means Georgian startups can build AI-powered products with minimal infrastructure investment.
On-device AI also solves the privacy equation — sensitive data never leaves the device, which is crucial for Georgian businesses handling personal data. Local AI also bypasses potential API cost barriers and latency issues from remote servers.
The Future of Mobile AI
Bonsai 27B represents a paradigm shift. AI models that were once only accessible via expensive cloud APIs can now run locally. This opens up categories like persistent on-device agents, offline-capable assistants, and privacy-preserving AI applications.