Kai-30B-Instruct

A 30B-parameter instruction-tuned language model optimized for reasoning, math, and code generation tasks, powered by our ADS (Adaptive Dual-Search Distillation) technique. The largest model in the Kai family.

Model Details

Model	Kai-30B-Instruct
Architecture	Qwen2ForCausalLM
Parameters	~30B
Hidden size	5120
Intermediate size	27648
Layers	64
Attention heads	40 (8 KV heads, GQA)
Context length	32768
Precision	bfloat16
Vocab size	152064
Chat template	ChatML (<|im_start|> / <|im_end|>)

Benchmark Results (5-shot, acc_norm)

Benchmark	Kai-30B-Instruct	Llama-3 70B	Qwen2.5 32B	Yi-34B	Llama-3 8B	Mistral 7B	Llama-2 7B
ARC-C	64.0	83.0	70.5	65.3	60.1	55.5	53.0
HellaSwag	74.4	89.0	85.2	83.1	78.6	81.3	78.6
PIQA	84.8	85.0	84.1	82.5	79.8	82.1	78.1
Winogrande	86.4	83.0	78.2	76.4	73.0	74.0	69.1

What is ADS?

Adaptive Dual-Search Distillation treats model fine-tuning as a constrained optimization problem inspired by Operations Research. The core mechanism is a dynamic loss function with a stateful dual penalty factor that adapts based on embedding space entropy — forcing the model to converge to high-confidence predictions at difficult reasoning points, without modifying the model architecture.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "NoesisLab/Kai-30B-Instruct",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("NoesisLab/Kai-30B-Instruct")

messages = [{"role": "user", "content": "What is 25 * 4?"}]
input_ids = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt"
).to(model.device)

output = model.generate(
    input_ids,
    max_new_tokens=512,
    temperature=0.6,
    top_p=0.8,
    do_sample=True,
)
print(tokenizer.decode(output[0][input_ids.shape[-1]:], skip_special_tokens=True))

Citation

@misc{noesislab2026kai30b,
  title={Kai-30B-Instruct},
  author={NoesisLab},
  year={2026},
  url={https://huggingface.co/NoesisLab/Kai-30B-Instruct}
}

License

Apache 2.0