Kiyo-65M

Kiyo-65M is a decoder-only language model pretrained from scratch on 80B tokens drawn from FineWeb-Edu, DCLM-Baseline, FineMath, and Stack-v3-train.

It is the smallest member of the Kiyo family and shares the design of Kiyo-135M: a Llama-style decoder with grouped query attention, RMSNorm, SwiGLU MLPs, and tied input/output embeddings, following the SmolLM2 architecture family and reusing its 49k-token vocabulary. Kiyo-65M scales that recipe down to 15 layers and a 512-dimensional residual stream, roughly halving the parameter count while keeping the same data mixture and training code. It is trained independently from a random initialization, not distilled or pruned from Kiyo-135M.

Model Details

Field	Value
Parameters	64,994,816
Architecture	Llama-style decoder (SmolLM2 architecture)
Layers	15
Hidden size	512
Intermediate size	1,216
Attention heads	8
KV heads	4
Head dimension	64
Attention type	Grouped query attention
Activation	SwiGLU
Normalization	RMSNorm
Positional encoding	RoPE (theta 10,000)
Vocabulary size	49,152
Context length	2,048
Embeddings	Tied input/output embeddings
Training tokens	79,999,533,056
Optimizer steps	152,587
Tokens per optimizer step	524,288
Peak learning rate	2.0e-3
LR schedule	Warmup-stable-decay (1/sqrt)
Weight format	safetensors

Training Data

Source	Domain	Mixture share
FineWeb-Edu	General web text, education-filtered	45.0%
DCLM-Baseline	General web text, high-quality filtered	30.0%
Stack-v3-train	Source code	17.5%
FineMath	Mathematical reasoning	7.5%

Benchmarks

Self-reported result from the official BananaMind Base Bench 1.1 script, measured on CUDA in float32.

Model	Params	Overall Elo
Kiyo-65M	65.0M	1,067
Rose-Mini	49.4M	1,037
BananaMind-2-Medium	49.6M	1,034
Supra-50M-Reasoning	51.8M	985

Figures for Rose-Mini, BananaMind-2-Medium, and Supra-50M-Reasoning are taken from the official BananaMindBench Leaderboard, all against the same BananaMind Base Bench 1.1 suite. Note that all three comparison models are 20–30% smaller than Kiyo-65M, so this is not a parameter-matched comparison.

Detailed Kiyo-65M result

Category	Accuracy	z vs. chance	Elo	Significant
Language completion	98.0%	+11.92	1,468	*
World knowledge	74.0%	+8.00	1,099	*
Commonsense	68.0%	+7.02	1,062	*
Code completion	62.0%	+6.04	1,193	*
Context tracking	48.0%	+3.76	949	*
Logical reasoning	42.0%	+2.78	1,007	*
Quantitative	34.0%	+1.47	918

* = passes 1.96σ vs. chance; n=50 per category

By difficulty

Difficulty	Accuracy
Easy	70.9%
Medium	59.8%
Hard	51.7%

Summary

Metric	Value
Parameters	64,994,816
Overall Elo	1,067
Chance floor	805
Above chance floor	+263
Raw accuracy	60.9%
95% CI on accuracy	55.7% – 66.0%

Scores are self-evaluated and may vary with the benchmark revision, Transformers version, dtype, hardware, and generation settings.

Usage

pip install -U transformers safetensors torch

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "DedeProGames/Kiyo-65M"

tokenizer = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16 if torch.cuda.is_available() and torch.cuda.is_bf16_supported() else torch.float16,
).cuda().eval()

prompt = "The meaning of life is "
input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(model.device)

with torch.no_grad():
    output = model.generate(
        input_ids=input_ids,
        max_new_tokens=64,
        do_sample=False,
        repetition_penalty=1.1,
        pad_token_id=tokenizer.eos_token_id,
        eos_token_id=tokenizer.eos_token_id,
    )

print(tokenizer.decode(output[0], skip_special_tokens=True))

Limitations

This is a base model, not instruction-tuned — it continues text rather than following instructions. At 65M parameters and 80B training tokens it still produces fluent, grammatical continuations and remains strong on language completion, but accuracy falls off sharply on quantitative reasoning, multi-step logic, and context tracking, where it sits close to chance on the harder items. The 2,048-token context window also limits long-document use. It can generate incorrect facts and should not be used for high-stakes decisions without verification. Keep a finite generation limit to avoid repetition or drift on long outputs.

License

Apache 2.0