Bonsai 27B (PrismML) Pricing
All plans and pricing as of 2026-07-18
Self-hosted (Free)
- ✓Apache 2.0 -- commercial use allowed
- ✓Ternary variant: 5.9GB ({-1,0,+1} weights, 1.71 effective bits/weight, ~95% of baseline quality)
- ✓1-bit variant: 3.9GB ({-1,+1} weights, 1.125 effective bits/weight, ~90% of baseline quality)
- ✓Native MLX (Mac/iPhone/iPad) + CUDA (NVIDIA) support
Locally AI iOS app
- ✓Ships inside the 'Locally AI (Local AI Chat)' iOS app -- on-device, no server round-trip
API preview
- ✓Free limited-time developer preview API
- ✓Also served via Together.ai
Is Bonsai 27B (PrismML) Worth the Price?
Value Score: 9.5/10
Overall Score: 7.9/10 · On-device AI builders and privacy-first users who want real reasoning, vision, and tool-calling on a phone or fanless laptop -- and local-AI hobbyists who want the best capability-per-gigabyte available.
Bonsai 27B is the most interesting local-AI release of the summer: it moves the 'runs on a phone' frontier from 3-8B toys to a genuine 27B-class model with vision, tool calls, and 262K context, and it publishes its own quality losses instead of hiding them. The honest framing matters -- you're getting ~90-95% of Qwen3.6-27B, which was already the open-weights coding champion in its size class, in 3.9-5.9GB. If you build on-device AI experiences or just want capable private inference in your pocket, this is the new reference point. If you have real hardware, the full-precision base model remains the better tool.
The Tier List Tuesday
Weekly newsletter: tier movers, new entrants, and the VS of the week. Built from our daily AI-tool sweeps. No spam, unsubscribe anytime.
How Bonsai 27B (PrismML) Pricing Compares
| Tool | Free Tier | Starting Price | Value Score | Overall |
|---|---|---|---|---|
| Bonsai 27B (PrismML)(this tool) | Yes | $0 | 9.5/10 | 7.9 |
| Qwen (Alibaba) | Yes | $0 | 10/10 | 8.8 |
| MiniMax M3 | Yes | $0 | 9.5/10 | 8.4 |
| Gemma 4 (Google) | Yes | $0 | 10/10 | 8.3 |
| IBM Granite 4.0 | Yes | $0 | 9.5/10 | 8.2 |
| Kimi K3 (Moonshot) | Yes | $3 / $15/per 1M tokens (input/output) | 8.5/10 | 8.1 |