A small Mixtral-style Mixture-of-Experts causal language model (~432M total / ~176M active parameters), trained from scratch on a single consumer GPU (RTX A5000 24GB).

Quick Start

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "mikecovlee/tinymixtral", trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("mikecovlee/tinymixtral")

Model Description

TinyMixtral is a research-oriented small MoE language model exploring how far data quality and training recipe can push a model at the ~432M parameter scale. It uses a Mixtral-style sparse MoE architecture with 6 experts and top-2 routing, achieving ~176M active parameters per forward pass.

This checkpoint (v1.1) is the result of a two-stage training pipeline:

Pretrain: 4B tokens on a SmolLM-inspired blend (FineWeb-Edu 89% + Cosmopedia v2 11%) with LR=7e-4
Post-train: 1B tokens on Wikipedia + Cosmopedia v2 (50:50) with LR=2e-5

The v1 model (C4 pretrain) is available at mikecovlee/tinymixtral-v1.0.

Training Details

Parameter	Pretrain	Post-train
Data	FineWeb-Edu + Cosmopedia v2 (89:11)	Wikipedia + Cosmopedia v2 (50:50)
Tokens	4B	1B
Steps	162,761	40,691
Learning rate	7e-4	2e-5
Warmup steps	2,000	300
Batch size	24 × 1024	24 × 1024
Optimizer	AdamW (β=0.9, 0.95, wd=0.1)	same
LR schedule	Cosine decay	Cosine decay
Precision	bf16	bf16
Training time	~84h (RTX A5000)	~20.5h
End loss	2.5	2.5

Evaluation Results

GLUE (zero-shot, 500 samples per task)

Task	Metric	v1 (C4)	v1.1
SST2	accuracy	0.470	0.568
MRPC	accuracy / F1	0.338 / 0.069	0.686 / 0.813
QQP	accuracy / F1	0.470 / 0.412	0.350 / 0.519
QNLI	accuracy	0.494	0.458
RTE	accuracy	0.520	0.534
MNLI	accuracy	0.348	0.352
MNLI-mm	accuracy	0.368	0.364
Mean		0.383	0.515

ARC

Task	v1 (C4)	v1.1 (0-shot / 5-shot)
ARC-Challenge	0.220 / 0.223	0.249 / 0.254
ARC-Easy	0.311 / 0.320	0.365 / 0.368

Key Findings

MRPC F1 improved from 0.069 (random) to 0.813 — data quality alone unlocked paraphrase understanding
GLUE Mean improved 34% (0.383 → 0.515) with zero architecture changes
Post-training with domain-specific data gave neutral results at this scale — the pretrain recipe is the dominant factor

Architecture

Parameter	Value
hidden_size	896
num_layers	10
Attention	GQA (14 Q heads / 2 KV heads)
Head dim	64
RoPE theta	1,000,000
Norm	RMSNorm
Experts	6 (top-2 routing, SwiGLU)
Vocab size	32,000 (SentencePiece)
Max position	2,048
Total params	~432M
Active params	~176M

Intended Use

This model is intended for research and educational purposes:

Studying MoE training dynamics at small scale
Experimenting with data quality ablations
Learning how to train language models from scratch on consumer hardware

This is a base language model (not instruction-tuned). It performs next-token prediction and is not designed for chat or instruction following.

Limitations

432M parameters limits factual knowledge and reasoning depth
Not instruction-tuned — will not follow chat-style prompts reliably
English-only training data
Random or near-random on several GLUE tasks (QNLI, QQP accuracy)

Citation

bibtex

@software{tinymixtral2026,
  author = {Michael Lee},
  title = {TinyMixtral: A Small Mixtral-Style MoE Language Model},
  year = {2026},
  url = {https://github.com/mikecovlee/tinymixtral}
}

mikecovlee/
tinymixtral

TinyMixtral v1.1