Ternary Bonsai 2 27B: 9x smaller model keeping 98.2% of benchmark scores

Melvin Vivas · X post · 2026-09-18 · Open on X

Topics: LLM Fundamentals, Industry Trends & Job Market · Level: intermediate

Summary

Ternary Bonsai 2 27B has been released. It is a ternary-quantized model based on Qwen3.8 27B. The quoted announcement says it is 9x smaller than the full-precision original while keeping 98.2% of its aggregate benchmark performance. It shows how heavy quantization can make large open models practical to run locally.

Key points

• Ternary Bonsai 2 27B is based on Qwen3.8 27B.
• Ternary (very low-bit) weights make it about 9x smaller than the full-precision model.
• It keeps 98.2% of the original's aggregate benchmark performance.
• It came out two months after the first Bonsai 27B, and the main improvement is quality.
• The creator is trying it out, so it's worth testing for local or low-cost inference.

Resources mentioned

• Ternary Bonsai 2 27B · tool · prismml.com · free
  A ternary-quantized 27B model, 9x smaller than full precision, that keeps 98.2% of benchmark performance.
• Qwen3.8-27B · tool · huggingface.co · free
  A 27B-parameter open-weight model from Alibaba's Qwen family that you can run locally or call through hosted APIs.
  Also in: Deploying Qwen3.8 27B on Hugging Face Inference Endpoints (Melvin Vivas on X · notes), Using Hugging Face credits: Jobs, Inference Endpoints and Open Models (Melvin Vivas on X · notes), Running Qwen3.8-27B Locally on an M5 Max MacBook with Inco Splash (Melvin Vivas on X · notes), Free Qwen-3.8 27B Model via Infron (Melvin Vivas on X · notes) and 24 more

Try this

• Try Ternary Bonsai 2 27B and compare it with the full-precision Qwen3.8 27B.