Ternary Bonsai 2 27B
Experimental SafeTensors checkpoint for the Ternary Bonsai 2 27B text model. This repository contains the converted model weights together with the tokenizer, chat template, and configuration required by a compatible runtime.
Model Details
- Architecture:
Bonsai2ForCausalLM - Configuration type:
qwen3_5_text - Quantization: Bonsai ternary (
quant_method: bonsai_ternary) - Weight format: SafeTensors
- Chat template: included as
chat_template.jinja
The quantization configuration uses an automatic ternary target and a normalized Sylvester-Walsh-Hadamard transform with 1024-element blocks and explicit signs. The checkpoint is not a conventional BF16 or FP16 model.
Intended Use
This checkpoint is intended for research, runtime integration, and evaluation
of ternary model inference. It requires a runtime that implements the
Bonsai2ForCausalLM architecture and the bonsai_ternary quantization format.
The corresponding Valendra vLLM implementation is available at:
https://github.com/valendra-tech/vllm/pull/3
Loading
Use a compatible Valendra vLLM build and pass this repository as the model identifier, for example:
vllm serve valendra/Ternary-Bonsai-2-27B-safetensors
Standard Transformers installations may not load this checkpoint unless they provide the Bonsai2 model implementation and quantization support.
Limitations
- This is an experimental converted and quantized artifact.
- No general accuracy, safety, or production-readiness claims are made here.
- Compatibility and performance depend on the runtime, GPU architecture, and the available CUDA kernels.
- Validate outputs and resource requirements before using the model in a production system.
No training-data, fine-tuning, or safety-evaluation information is asserted by this model card. Review the license and usage terms of the upstream model before redistributing or deploying this artifact.