nanowhale-100m-base 🐳

A small ~110M parameter language model implementing the DeepSeek-V4 architecture from scratch. This is the pretrained base model — see cmpatino/nanowhale-100m for the SFT/chat version.

Training code: github.com/huggingface/nanowhale

Architecture

This model implements key DeepSeek-V4 innovations at a miniature scale:

Component	Details
Parameters	~110M total (41M embeddings, 69M non-embedding)
Hidden size	320
Layers	8
Attention heads	8 (1 KV head — MQA-style)
Head dim	96 (32 RoPE + 64 NoPE)
MLA	q_lora_rank=160, o_groups=2, o_lora_rank=80
MoE	4 routed experts + 1 shared, top-2 routing
Expert FFN	SwiGLU, intermediate_size=640
Routing	sqrtsoftplus scoring, noaux_tc method
Hyper-Connections	hc_mult=4, Sinkhorn routing (2 iters)
MTP	1 next-token prediction layer
Vocab	129,280 (DeepSeek-V4 tokenizer)
Context	2,048 tokens

DeepSeek-V4 Features Implemented

• Multi-head Latent Attention (MLA): Compressed KV cache via latent projections
• Mixture of Experts (MoE): Sparse activation — only 2 of 4 experts per token
• Hyper-Connections: Multi-copy hidden states with learned Sinkhorn routing replacing residual connections
• SwiGLU FFN with configurable limit
• Grouped output projection (o_groups)

Training

• Dataset: HuggingFaceFW/fineweb-edu (streaming)
• Steps: 5,000
• Tokens seen: ~2.6B
• Batch size: 8 × 4 gradient accumulation = 32 effective
• Sequence length: 2,048
• Learning rate: 6e-4, cosine schedule, 3% warmup
• Optimizer: AdamW (β1=0.9, β2=0.95, weight_decay=0.1)
• Precision: bf16 mixed precision
• Hardware: 1× NVIDIA H100 80GB

Training Metrics

Metric	Value
Final loss	~5.3 (cross-entropy)
Final entropy	3.77
Token accuracy	33.8%

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "cmpatino/nanowhale-100m-base", trust_remote_code=True, dtype=torch.float32
).cuda().eval()
tokenizer = AutoTokenizer.from_pretrained("cmpatino/nanowhale-100m-base")

input_ids = tokenizer.encode("The meaning of life is", return_tensors="pt").cuda()
output = model.generate(input_ids, max_new_tokens=100, temperature=0.7, top_p=0.9,
                        pad_token_id=tokenizer.eos_token_id)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Limitations

• Small model: 110M params with 129K vocab means ~37% of parameters are in embeddings, limiting model capacity
• Limited training: Only 5K steps / 2.6B tokens — significantly undertrained compared to production models
• Pretrained only: This is a base model without instruction tuning. Outputs are language-model completions, not conversations.
• fp32 recommended: The Hyper-Connections architecture can produce values that overflow bf16 range at this scale. Use dtype=torch.float32.
• Custom architecture: Requires trust_remote_code=True

License

Apache-2.0