kshitijthakkar/deepseek-v4-mini-3B-init

DeepSeek-V4 Mini (3B) — randomly-initialized architecture replica

Faithful small-scale (~3.2B total / ~1.10B activated per token) replica of the DeepSeek-V4 architecture, sized to be trainable on rented GPUs and to map cleanly onto the full-scale V4-Flash dimensions for weight slicing.

This is a randomly-initialized scaffold — generates noise. Its purpose:

• reference architecture for ablation / hyperparameter-search experiments
• target for weight transfer / slicing from real V4-Pro / V4-Flash

Architecture summary

	Value
hidden_size	1536
num_hidden_layers	28
num_attention_heads	24
num_key_value_heads	1 (MQA)
head_dim	64
q_lora_rank / o_lora_rank	512 / 512
qk_rope_head_dim	32
o_groups	4
n_routed_experts	24
n_shared_experts	1
num_experts_per_tok	4
num_hash_layers	2
moe_intermediate_size	768
compress_ratios	[0, 0, 4, 112, 4, 112, 4, 112, 4, 112, 4, 112, 4, 112, 4, 112, 4, 112, 4, 112, 4, 112, 4, 112, 4, 112, 4, 0]
index_topk / heads / head_dim	192 / 16 / 96
sliding_window	64
max_position_embeddings	1,048,576 (YaRN factor=16)
vocab_size	129280 (real V4-Flash tokenizer)
num_nextn_predict_layers	1 (V3-style MTP)
hc_mult (n_hc)	4
Storage dtype	bfloat16

Quick start

from huggingface_hub import login, snapshot_download
login()  # private repo

local = snapshot_download(repo_id="kshitijthakkar/deepseek-v4-mini-3B-init")

import sys, os
sys.path.insert(0, os.path.join(local, "code"))
import deepseek_v4  # registers DeepseekV4{Config,ForCausalLM} with HF auto classes

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained(local)
model = AutoModelForCausalLM.from_pretrained(local, torch_dtype=torch.bfloat16)
model.eval()

ids = tok.apply_chat_template(
    [{"role": "user", "content": "Hello"}],
    return_tensors="pt", add_generation_prompt=True, return_dict=True,
)
with torch.no_grad():
    out = model(input_ids=ids["input_ids"])
print(out.logits.shape)

Components implemented

mHC (Sinkhorn-Knopp) · CSA + Lightning Indexer · HCA · pure sliding-window · Shared-KV MQA + grouped output projection (per-group wo_a) · partial RoPE + output -i rotation · attention sink · DeepseekMoE with sqrt(softplus) routing · hash-routed early layers · clamped SwiGLU · MTP head · YaRN.

Every component is bit-equivalent in math to the official inference/model.py

• kernel.py:hc_split_sinkhorn (FP4/FP8 quantization and Hadamard rotation are skipped — those are inference optimizations, not architecture).

Citation

@misc{deepseek_v4_2026,
  author = {DeepSeek-AI},
  title  = {DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence},
  year   = {2026},
  url    = {https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash}
}