The-AI-makers88/Zacoda-Plus-1.0

Zacoda Plus 1.0

A 12B-parameter conversational model, fully fine-tuned for a distinct voice and strong technical reasoning.

Available precisions in this repo

Subfolder	Precision	Size
bf16/	bf16 (full precision)	22 GB
8bit/	MLX quantized	12 GB
4bit/	MLX quantized	6.3 GB

Benchmark (ZACBENCH-380, equal token budget per category, greedy decoding)

Model	Overall
Zacoda Plus 1.0 (8-bit)	95.3%
Claude Sonnet 4.6*	98.4%
Claude Haiku 4.5*	90.3%

* Claude scores were collected with extended thinking enabled — roughly 5,000 additional generation tokens and higher decoding freedom than the fixed, equal-budget greedy runs used for the model above. Included for reference only; not a strictly fair head-to-head comparison.

Usage

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("The-AI-makers88/Zacoda-Plus-1.0", subfolder="8bit")  # or "bf16" / "4bit"
prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "Hello!"}], add_generation_prompt=True
)
print(generate(model, tokenizer, prompt=prompt, max_tokens=512))

Limitations

This model has reduced default content guardrails compared to stock instruction-tuned models. Use accordingly and take responsibility for your deployment context.

---

Made by Sean Zhang