Instructions to use Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF:IQ2_XXS
# Run inference directly in the terminal:
llama cli -hf Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF:IQ2_XXS

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF:IQ2_XXS
# Run inference directly in the terminal:
llama cli -hf Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF:IQ2_XXS

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF:IQ2_XXS
# Run inference directly in the terminal:
./llama-cli -hf Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF:IQ2_XXS

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF:IQ2_XXS
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF:IQ2_XXS

Use Docker

docker model run hf.co/Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF:IQ2_XXS

LM Studio
Jan

vLLM

How to use Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF:IQ2_XXS

Ollama
How to use Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF with Ollama:
```
ollama run hf.co/Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF:IQ2_XXS
```
Unsloth Desktop

How to use Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF:IQ2_XXS

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF:IQ2_XXS"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF with Docker Model Runner:
```
docker model run hf.co/Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF:IQ2_XXS
```

Lemonade

How to use Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF:IQ2_XXS

Run and chat with the model

lemonade run user.DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF-IQ2_XXS

List all available models

lemonade list

Hermes Agent

How to use Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF:IQ2_XXS

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF:IQ2_XXS

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF:IQ2_XXS

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF:IQ2_XXS" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

DeepSeek V4 Flash 0731 — L37-42 Q4_K, Joint-Finetuned MTP (GGUF)

A quantized GGUF of DeepSeek V4 Flash (0731 checkpoint) with a jointly finetuned inline Multi-Token Prediction (MTP) drafter. This is the 0731 revision of DeepSeek V4 Flash. Unlike the released Flash GGUFs, this build embeds the 3-stage MTP drafter directly in the main model — no separate support/sidecar file is required. It is intended for the DwarfStar inference engine, which is the only loader that understands this layout.

File name (HF convention): DeepSeek-V4-Flash-IQ2_XXS-L37-42Q4K-MTP-JointFT-imatrix-0731.gguf

Fine-tuning data

The joint fine-tuning dataset is derived from questions posed to Fable 5, Opus 4.8, and Opus 4.7. Coding repositories from GitHub were part of the MTP drafter re-training corpus, not of this joint fine-tuning dataset.

Model details


Architecture	DeepSeek V4 Flash (`deepseek4`), 43 layers
Base	`deepseek-ai/DeepSeek-V4-Flash`, 0731 checkpoint
Revision	0731
Parameters	284B total / 13B active (base); ~304B logical incl. MTP stages
Context length	1,048,576 tokens (train)
Attention	64 heads, 1 KV head, head_dim 512, sliding window 128
Experts	256 routed, 6 used per token
MTP drafter	3 inline stages, block size 5, target layers 40–42
File size	96.47 GiB
License	MIT (base DeepSeek-V4-Flash is MIT)

Quantization recipe

Asymmetric mixed quantization: only the routed MoE experts are aggressively compressed; projections, routing, and output are kept high-precision to preserve quality. imatrix-tuned.

Component	Type
Layers 37–42 routed experts	`Q4_K` (closest to output, higher quality)
Layers 0–36 routed gate/up experts	`IQ2_XXS`
Layers 0–36 routed down experts	`Q2_K`
Attention projections	`Q8_0`
Shared experts	`Q8_0`
Output head	`Q8_0`
Norms / HC / compressor / indexer	`F16` / `F32`

Tensor-type breakdown (from ds4 --inspect):

Type	Tensors	Size
`iq2_xxs`	80	41.25 GiB
`q2_k`	40	26.25 GiB
`q4_k`	18	20.25 GiB
`q8_0`	376	6.66 GiB
`f16`	366	2.05 GiB
`f32`	527	~0 GiB
`i32`	3	~0 GiB

Inline MTP drafter (the distinguishing feature)

The 3-stage MTP drafter is embedded in the main GGUF and jointly finetuned with the main model (mtp/dspark metadata: stages=3, block=5, target_layers=40,41,42). Because the drafter ships inside the model, you do not need a separate --mtp support GGUF. DwarfStar binds it automatically at load:

ds4: embedded_mtp stages=3 bound_stages=3 source=main_model block_size=5 draft=1

Current status of MTP drafting on this build: speculative decoding is available via --mtp-draft N but is presently net-negative on this finetuned checkpoint — verification costs more than accepted drafts save (measured ~31 t/s with draft-2 vs ~37 t/s plain at 13k context, byte-identical output). The engine therefore defaults to --mtp-draft 1 (plain decoding). Drafting remains opt-in for experimentation as drafter acceptance improves.

Intended use

Local inference of DeepSeek V4 Flash on high-memory Apple Silicon (Metal), NVIDIA CUDA, or ROCm, via the DwarfStar engine.
Fits 128 GB unified-memory machines (96.47 GiB weights + KV/scratch).
Research into inline/joint MTP speculative decoding.

Out of scope: use with generic GGUF loaders (llama.cpp etc.) — this file requires the DwarfStar engine's tensor layout and inline-MTP handling.

Performance (DwarfStar, Metal)

Single-run, --ctx 32768, greedy, Mac Studio M3 Ultra 512 GB:

Context	Prefill	Generation (default)	Generation (draft-2)
~13k tokens	~470 t/s	~37 t/s	~31 t/s

Warm KV-cache reuse (disk checkpoint, alignment off) reduces a repeat 18k-turn prefill from ~36.8 s to ~3.4 s.

How to run

# Download (once available in antirez/deepseek-v4-gguf)
./download_model.sh q2-q4-mtp-imatrix

# Plain decoding (default, currently fastest)
./ds4 -m gguf/DeepSeek-V4-Flash-IQ2_XXS-L37-42Q4K-MTP-JointFT-imatrix-0731.gguf --temp 0

# Experiment with MTP drafting (currently net-negative, opt-in)
./ds4 -m gguf/DeepSeek-V4-Flash-IQ2_XXS-L37-42Q4K-MTP-JointFT-imatrix-0731.gguf \
      --mtp-draft 2 --temp 0

Limitations & biases

Inherits DeepSeek V4 Flash's limitations and biases. Quantization is lossy; the 2-bit routed-expert compression is imatrix-tuned to preserve quality but is not bit-exact with the FP8 base. MTP drafting is experimental and currently slower than plain decoding on this checkpoint.

License

MIT. Derived from deepseek-ai/DeepSeek-V4-Flash (MIT).

Citation

@misc{deepseekai2026deepseekv4,
      title={DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence},
      author={DeepSeek-AI},
      year={2026},
      url={https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash}
}

Downloads last month: 95

GGUF

Model size

304B params

Architecture

deepseek4

Hardware compatibility

2-bit

Model tree for Deviad/DeepSeek-V4-Flash-0731-MTP-JointFT-GGUF

Base model

deepseek-ai/DeepSeek-V4-Flash

Quantized

(127)

this model