Experimenting with Ternary Model Fine-tuning for Agentic Coding

professormunchies · reddit · 2026-07-19

The author attempts post-training fine-tuning on true ternary/binary low-bit models like PrismML Bonsai 8B, aiming to make it function more like an agentic coding model, and tests it on 10 Python issues from SWE-rebench-v1.

Key Background

Bonsai is not pre-trained from scratch using the BitNet approach; instead, it converts existing Qwen3 models into ternary/binary weights.
The ternary model is approximately 1.7 bits/weight, meaning the 8B version takes up only about 2GB.
This is not a standard GGUF quantization. Calibration methods for quantization error like imatrix/AWQ/GPTQ are essentially ineffective because the checkpoint itself is scale × {-1,0,+1}.

Fine-tuning Method

Continued training uses standard QAT: weights are ternarized during the forward pass, and a straight-through estimator is used for the backward pass.
Training was done on Metal without relying on CUDA.
After training, it was re-ternarized and packed back into a 2-bit GGUF, maintaining its original size.

Results

|---|---:|---:|---:|---|

| base 8B (unfine-tuned) | 50% | 0% | - | - |

| QAT, trained only last 18 layers | 40% | 0% | 1.0 | Falls into repetitive loops |

| QAT, trained by "gradient impact layers" | 40% | 0% | 1.0 | Loop issue fixed, runs more stable |

| QAT, trained all 36 layers | 30% | 0% | 0.91 | Most stable, but worst patch rate |

Conclusion Leanings

A drop in training loss does not mean the model has actually learned coding abilities.
The author emphasizes the need to distinguish whether "the weight code actually flipped" or if "only the scale changed". Otherwise, it may look like it learned something while the underlying logic remains unchanged.

Original post →

Experimenting with Ternary Model Fine-tuning for Agentic Coding

More from coding & agent