
Model	Kai-0.35B-Instruct
Architecture	LlamaForCausalLM
Parameters	360M
Hidden size	960
Layers	32
Attention heads	15 (5 KV heads, GQA)
Context length	8192
Precision	bfloat16
Vocab size	49,152

Benchmark Results (5-shot, log-likelihood)

Benchmark	Kai-0.35B-Instruct	Mamba (370M)	TinyLlama (1.1B)	Llama-3.2 (1B)
ARC-Challenge (science reasoning)	37.80%	~29.1%	~30.1%	~44.5%
HellaSwag (sentence completion)	55.88%	~53.8%	~59.2%	~61.1%
PIQA (physical commonsense)	71.82%	~69.6%	~73.0%	~74.5%

Code Generation — MBPP (3-shot, pass@1)

Model	Params	MBPP pass@1
Mamba / Mamba-2	370M	<10.0%
TinyLlama	1.1B	~19.91%
Kai-0.35B-Instruct	360M	22.20%
Llama-3.2-1B (Base)	1.0B	~25-30%
Llama-3.2-1B-Instruct	1.0B	~49.0%

Key Observations

ARC-Challenge: Kai-0.35B scores 37.80% (5-shot), significantly outperforming both Mamba-370M (+8.7pp) and TinyLlama-1.1B (+7.7pp) — a model 3x its size.
PIQA: At 71.82%, Kai-0.35B nearly matches TinyLlama-1.1B (73.0%) with only 1/3 the parameters, and trails the 1B-class Llama-3.2 by less than 3pp.
MBPP: At 22.20% pass@1, Kai-0.35B surpasses TinyLlama-1.1B (~19.91%) in code generation despite being 3x smaller.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model = AutoModelForCausalLM.from_pretrained(
    "NoesisLab/Kai-0.35B-Instruct",
    torch_dtype=torch.bfloat16,
)
tokenizer = AutoTokenizer.from_pretrained("NoesisLab/Kai-0.35B-Instruct")
messages = [{"role": "user", "content": "What is 25 * 4?"}]
input_ids = tokenizer.apply_chat_template(messages, return_tensors="pt")
output = model.generate(input_ids, max_new_tokens=256)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Citation

bibtex

@misc{noesislab2026nkai,
  title={Kai-0.35B-Instruct},
  author={NoesisLab},
  year={2026},
  url={https://huggingface.co/NoesisLab/Kai-0.35B-Instruct}
}

License

Apache 2.0