Instructions to use kshitijthakkar/deepseek-v4-mini-300M-from-flash with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use kshitijthakkar/deepseek-v4-mini-300M-from-flash with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="kshitijthakkar/deepseek-v4-mini-300M-from-flash")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("kshitijthakkar/deepseek-v4-mini-300M-from-flash")
model = AutoModelForCausalLM.from_pretrained("kshitijthakkar/deepseek-v4-mini-300M-from-flash")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use kshitijthakkar/deepseek-v4-mini-300M-from-flash with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "kshitijthakkar/deepseek-v4-mini-300M-from-flash"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "kshitijthakkar/deepseek-v4-mini-300M-from-flash",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/kshitijthakkar/deepseek-v4-mini-300M-from-flash

SGLang

How to use kshitijthakkar/deepseek-v4-mini-300M-from-flash with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "kshitijthakkar/deepseek-v4-mini-300M-from-flash" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "kshitijthakkar/deepseek-v4-mini-300M-from-flash",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "kshitijthakkar/deepseek-v4-mini-300M-from-flash" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "kshitijthakkar/deepseek-v4-mini-300M-from-flash",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use kshitijthakkar/deepseek-v4-mini-300M-from-flash with Docker Model Runner:
```
docker model run hf.co/kshitijthakkar/deepseek-v4-mini-300M-from-flash
```

DeepSeek-V4 Mini (300M) — sliced from V4-Flash

A weight-transferred small replica of the DeepSeek-V4 architecture. Tensors were sliced (or, where shapes already matched, copied verbatim) from deepseek-ai/DeepSeek-V4-Flash into a smaller architecture matching kshitijthakkar/deepseek-v4-mini-300M-init.

This is not a randomly-initialized scaffold — most parameters carry real training signal from the source. It is intended as a warm-start for fine tuning at this size, not as a finished chat / instruct model. See the fine-tune recipe below.

	Value
Total params	~0.32 B
Activated per token	~0.17 B
Source	deepseek-ai/DeepSeek-V4-Flash (284 B total / 13 B active)
Random-init baseline	kshitijthakkar/deepseek-v4-mini-300M-init
Vocab	129,280 (V4-Flash tokenizer + ChatML chat template baked in)
Storage dtype	bfloat16, 2 GB shards

Slice provenance

Each parameter in this model is one of three kinds:

Kind	Meaning	Recommended at fine-tune
COPY	Source dim already matched the destination — bytes copied verbatim.	Freeze (no training)
SLICE	Destination dim is smaller — deterministic slice (truncate / per-head / per-group / top-K experts by gate-norm).	Train (lossy slice needs to recover)
INIT	Source has no compatible counterpart (or shapes incompatible) — kept at random init.	Train (full learning)

Kind	Tensors	% of params
COPY	56	6.1%
SLICE	789	86.0%
INIT	72	7.9%

Concretely:

Layer pruning: source had 43 layers; we kept 12 via uniform stride: [0, 3, 7, 10, 14, 17, 21, 25, 28, 32, 35, 39]. The CSA/HCA/SW per-layer pattern is preserved.
Expert pruning: source had 256 routed experts per layer; we kept the top-16 by gate.weight L2 norm (a cheap proxy for expert importance — replace with an activation-statistics-based pick if you have profiling data).
Compressors / Indexer: marked INIT in this release (the per-block positional bias and gate require a non-trivial cross-dim mapping that the current slicer leaves at random; their parent attn KV/Q paths are sliced from source).
MTP module: sliced from the source's first MTP step using the same per-tensor rules.
Hash-routing tid2eid table: regenerated (top-k differs from source).

The exhaustive list of COPY (frozen-recommended) keys is in frozen_keys.json.

Quick start (Colab / local)

# 1) install + auth (private repo)
# !pip install -q "transformers>=4.50" huggingface_hub safetensors
from huggingface_hub import login, snapshot_download
login()

# 2) download repo (weights + tokenizer + modeling package)
local = snapshot_download(repo_id="kshitijthakkar/deepseek-v4-mini-300M-from-flash")

# 3) register the deepseek_v4 model_type with HF auto classes
import sys, os
sys.path.insert(0, os.path.join(local, "code"))
import deepseek_v4  # noqa: F401 — side-effect: AutoConfig / AutoModelForCausalLM registration

# 4) load via the standard Auto API
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

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

# 5) chat-templated forward pass — initial loss is much lower than the
# random-init scaffold, but the compressor/indexer paths are still random
# until you fine-tune (see below).
messages = [{"role": "user", "content": "Hello"}]
ids = tok.apply_chat_template(messages, return_tensors="pt",
                              add_generation_prompt=True, return_dict=True)
with torch.no_grad():
    print(model(input_ids=ids["input_ids"]).logits.shape)

Fine-tune recipe (warm-start)

Because most non-MoE-expert weights are already well-trained, freeze them and train only the lossy / random-init subset:

from training.warm_start import apply_warm_start_freeze
import os, json

frozen_path = os.path.join(local, "frozen_keys.json")
summary = apply_warm_start_freeze(model, frozen_path)
print(summary)
# -> trainable=Y M  frozen=X M  total=N M  train_ratio=...%

# Build optimizer over the trainable subset (configure_optimizers already
# respects requires_grad)
from training.optimizer import configure_optimizers
optimizer = configure_optimizers(
    model, learning_rate=2e-4, weight_decay=0.1,
    betas=(0.9, 0.95), device_type="cuda", optimizer_type="muon",
)

# ... standard training loop ...

Suggested starting hyperparameters:

	Value
LR (peak)	2 × 10⁻⁴ for trainable subset
Optimizer	Muon (2D weights) + AdamW (norms / 1-D)
Schedule	Linear warmup 200 steps → cosine to 10%
Sequence length	4096 (extend later via YaRN)
Batch tokens / step	0.5–2 M depending on hardware
MTP loss weight	0.3 (drops to 0.1 at end-of-decay)

Architecture summary

Parameter	Value
hidden_size	512
num_hidden_layers	12
num_attention_heads	8
num_key_value_heads	1 (MQA)
head_dim	64
q_lora_rank / o_lora_rank	256 / 256
qk_rope_head_dim	32
o_groups	2
n_routed_experts	16
n_shared_experts	1
num_experts_per_tok	2
moe_intermediate_size	512
num_hash_layers	2
sliding_window	32
max_position_embeddings	1,048,576 (YaRN factor=16, original=65536)
num_nextn_predict_layers	1 (MTP)
hc_mult (n_hc)	4

Limitations

Compressor + Indexer modules are random-init (see Slice provenance above). The current slicer doesn't do the cross-dim mapping for those per-block tensors. Forward-pass logits are noisy through CSA/HCA layers until those are fine-tuned.
Hash-table tid2eid regenerated — token routing for the first num_hash_layers differs from source.
No FP4/FP8 weights — source is dequantized to BF16 at slice time, so this repo doesn't carry V4-Flash's quantization-aware training state.

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}
}

paper PDF