Ternary Bonsai 2 27B: 5.9GB quantized model keeps 98.2% performance, runs on Mac
soumitrashukla9 · x · 2026-09-22
PrismML released Ternary Bonsai 2 27B, a ternary-quantized model based on Qwen3 27B that is 9x smaller than its full-precision counterpart, fits in 5.9GB, and retains 98.2% of aggregate benchmark performance, with notable gains in agentic coding, multimodal reasoning and long-horizon tool use; Apache 2.0 licensed. Economist johnjhorton then ran it locally on a 48GB M4 Max MacBook Pro via EDSL and a Metal-enabled llama.cpp fork, executing a battery of economic rationality tests successfully, with code and a write-up open-sourced.
Related event: PrismML Open-Sources Ternary Bonsai 2 27B: 9x Smaller, 98.2% Performance(3 posts)→
More from Models
- Average users aren't throwing frontier models at open math problems, dev observes — felpix_ · 2026-09-22
- Zero-day hits Meta's Muse for Mac: local process can steal prompts, auth tokens and file access — MicahBerkley · 2026-09-22
- Grok 4.7 spotted in user chatter as users hope for usage reset — BWay124 · 2026-09-22
- Amassing a PhD team is exactly what OpenAI did, dev notes in AI research debate — felpix_ · 2026-09-22
- Dev argues prompting alone can't get AI to solve natural science problems — felpix_ · 2026-09-22
- Benchmarks are meaningless, the differences between models are just vibes, dev argues — gnukeith · 2026-09-22