This experimental base model uses the exact BananaMind 2 Nano architecture and tokenizer. It was trained from scratch with stock PyTorch Muon on the hidden matrices and AdamW on the tied embedding and normalization weights, using only streamed FineWeb-Edu data for 24,999,591,936 custom-tokenizer tokens.

Architecture

Field	Value
Parameters	9,968,128
Layers	10
Hidden size	256
Intermediate size	768
Query heads	4
KV heads	2
Head dimension	64
Context	4,096
Vocabulary	8,192
Embeddings	Tied
Attention	GQA, pre-RoPE QK norm
MLP	SwiGLU
Position encoding	RoPE, theta 100,000

Training

Field	Value
Dataset	`HuggingFaceFW/fineweb-edu` / `sample-100BT`
Dataset revision	`87f09149ef4734204d70ed1d046ddc9ca3f2b8f9`
Data access	Streaming
Hidden-matrix optimizer	PyTorch Muon (`adjust_lr_fn="original"`)
Muon peak learning rate	0.05
Muon momentum	0.95, Nesterov
Muon Newton-Schulz steps	5
Embedding/norm optimizer	AdamW
AdamW peak learning rate	0.003
AdamW betas	(0.9, 0.95)
Global batch	132 sequences
Tokens per optimizer step	540,672
Optimizer steps	46,238
Warmup	1,750 steps
Schedule	Warmup-stable-decay, final 15% cosine cooldown
Weight decay	0.1, then 0.01 after 12,000,000,000 tokens
Precision	bfloat16 autocast, float32 master weights
Hardware	8 x NVIDIA RTX PRO 6000 Blackwell Server Edition
Seed	1337

The original Nano effective batch was 12 micro-batches x 11 accumulation steps = 132 sequences. This distributed run preserves that exact global batch. Ranks receive 16 or 17 sequences and scale their local mean losses so DDP's averaged gradient is the true 132-sequence global mean.

Usage

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Banaxi-Tech/muon-model-test"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    dtype=torch.bfloat16,
    device_map="auto",
)

This is a base model, not an instruction-tuned chat model.