Kai-3B-Instruct
A 3B-parameter instruction-tuned language model optimized for reasoning, math, and code generation tasks, powered by our new ADS (Adaptive Dual-Search Distillation) technique.
Model Details
| Model | Kai-3B-Instruct |
| Architecture | SmolLM3ForCausalLM |
| Parameters | 3B |
| Hidden size | 2048 |
| Intermediate size | 11008 |
| Layers | 36 |
| Attention heads | 16 (4 KV heads, GQA) |
| Context length | 65536 |
| Precision | bfloat16 |
| Vocab size | 128,256 |
What is ADS?
Adaptive Dual-Search Distillation (自适应对偶搜索蒸馏) 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.
Benchmark Results
General (5-shot, log-likelihood)
| Model | Params | MMLU | ARC-c (acc_norm) | HellaSwag (acc_norm) | PIQA (acc_norm) |
|---|---|---|---|---|---|
| TinyLlama | 1.1B | ~26.0% | ~33.0% | ~60.0% | ~71.0% |
| SmolLM2 | 1.7B | ~35.0% | ~38.0% | ~65.0% | ~74.0% |
| Llama-2-7B | 7B | 45.3% | 46.2% | 77.2% | 79.8% |
| Gemma-2-2B | 2.6B | ~52.0% | ~53.0% | 75.0% | ~78.0% |
| Kai-3B-Instruct | 3B | 53.62% | 51.88% | 69.53% | 77.53% |
| Qwen2.5-3B | 3B | ~63.0% | ~55.0% | ~73.0% | ~80.0% |
Code Generation — HumanEval (Pass@1, 0-shot)
| Model | Params | HumanEval (Pass@1) | Notes |
|---|---|---|---|
| Llama-2-7B | 7B | ~12.8% | 3x overtake — smaller model, far better code |
| SmolLM2-1.7B | 1.7B | ~25.0% | ADS delivers +14pp pure gain |
| Gemma-2-2B | 2B | ~30.0% | Surpasses Google's heavily distilled 2B flagship |
| Kai-3B-Instruct | 3B | 39.02% | ADS topological pruning, full pipeline |
| GPT-3.5 (Legacy) | 175B | ~48.0% | Kai-3B trails the original GPT-3.5 by only ~9pp |
Math — GSM8K (0-shot)
| Model | Params | GSM8K (exact_match) |
|---|---|---|
| Kai-3B-Instruct | 3B | 39.27% |
Key Observations
-
Surpasses Llama-2-7B: Kai-3B outperforms Llama-2-7B on MMLU (+8.3pp) and ARC-Challenge (+5.7pp) with less than half the parameters — a 7B model decisively beaten by a 3B distilled model.
-
Competitive with Gemma-2-2B: Matches or exceeds Google's Gemma-2-2B on MMLU (+1.6pp) and PIQA, despite Gemma being trained with significantly more compute.
-
HellaSwag: At 69.53%, Kai-3B surpasses all sub-2B models by a wide margin and trails the compute-heavy Qwen2.5-3B by only ~3.5pp.
-
PIQA: At 77.53%, Kai-3B nearly matches Gemma-2-2B (~78.0%) and approaches the 3B-class ceiling set by Qwen2.5-3B (~80.0%).
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model = AutoModelForCausalLM.from_pretrained(
"NoesisLab/Kai-3B-Instruct",
torch_dtype=torch.bfloat16,
)
tokenizer = AutoTokenizer.from_pretrained("NoesisLab/Kai-3B-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
@misc{noesislab2026kai3b,
title={Kai-3B-Instruct},
author={NoesisLab},
year={2026},
url={https://huggingface.co/NoesisLab/Kai-3B-Instruct}
}
License
Apache 2.0