A small ~110M parameter language model implementing the DeepSeek-V4 architecture, fine-tuned for chat/instruction following. Trained from scratch — no weights from DeepSeek-V4 were used.

Pretrained base model: cmpatino/nanowhale-100m-base
This model: SFT on HuggingFaceTB/smol-smoltalk
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)
MLA	Multi-head Latent Attention with q_lora_rank=160
MoE	4 routed experts + 1 shared, top-2 routing
Hyper-Connections	hc_mult=4, Sinkhorn routing (replacing residual connections)
MTP	1 next-token prediction layer
Vocab	129,280 (DeepSeek-V4 tokenizer)
Context	2,048 tokens

Training

Stage 1: Pretraining

Dataset: HuggingFaceFW/fineweb-edu
Steps: 5,000 | Tokens: ~2.6B
Batch: 32 effective (8 × 4 GA) | Seq length: 2,048
LR: 6e-4, cosine, 3% warmup
Precision: bf16 mixed

Stage 2: SFT (this model)

Dataset: HuggingFaceTB/smol-smoltalk (460K conversations)
Steps: 3,000 | Tokens: ~72.7M
Batch: 32 effective (8 × 4 GA) | Seq length: 2,048
LR: 2e-5, cosine, 5% warmup
Precision: fp32

Metrics

Metric	Pretrained	SFT
Eval loss	—	2.607
Perplexity (held-out)	13.62	12.90
Token accuracy	33.8%	48.5%

Usage

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

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

messages = [{"role": "user", "content": "What are 3 benefits of exercise?"}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
input_ids = tokenizer.encode(prompt, return_tensors="pt").cuda()
output = model.generate(input_ids, max_new_tokens=200, temperature=0.7, top_p=0.9,
                        pad_token_id=tokenizer.eos_token_id)
print(tokenizer.decode(output[0][input_ids.shape[1]:], skip_special_tokens=True))

Limitations

Tiny model: 110M params with 129K vocabulary — most capacity goes to embeddings. Generations are often incoherent or factually wrong.
Undertrained: Only 5K pretrain + 3K SFT steps. Production models train for 100K+ steps on trillions of tokens.
Educational purpose: This model demonstrates the DeepSeek-V4 architecture at small scale. It is not suitable for any production use.
fp32 recommended: The Hyper-Connections architecture can produce values that overflow bf16 range at this scale. Use dtype=torch.float32.
Custom code: Requires trust_remote_code=True.

Hardware

Trained on 1× NVIDIA H100 80GB.

License

Apache-2.0