DeepSeek-V4-Flash by DeepSeek: Open-Weight Model

Runs On

What it takes to serve DeepSeek-V4-Flash (290.9B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.

Precision	Weights	Memory needed	Cheapest setup	Per hour	Also fits
16-bit	581.9 GB	698.3 GB	3x MI325X (256 GB) Vultr	$6.00	4x MI300X $7.40 · 3x MI355X $7.77
8-bit	290.9 GB	349.1 GB	2x MI300X (192 GB) Vultr	$3.70	2x MI325X $4.00 · 2x MI355X $5.18
4-bit	145.5 GB	174.6 GB	1x MI300X (192 GB) Vultr	$1.85	1x MI325X $2.00 · 1x MI355X $2.59

Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Sep 18, 2026.

Model Card

By DeepSeek, published under mit, revision 60d8d70770c6.

DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence

Technical Report

Introduction

We present a preview version of DeepSeek-V4 series, including two strong Mixture-of-Experts (MoE) language models — DeepSeek-V4-Pro with 1.6T parameters (49B activated) and DeepSeek-V4-Flash with 284B parameters (13B activated) — both supporting a context length of one million tokens.

DeepSeek-V4 series incorporate several key upgrades in architecture and optimization:

Hybrid Attention Architecture: We design a hybrid attention mechanism combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) to dramatically improve long-context efficiency. In the 1M-token context setting, DeepSeek-V4-Pro requires only 27% of single-token inference FLOPs and 10% of KV cache compared with DeepSeek-V3.2.
Manifold-Constrained Hyper-Connections (mHC): We incorporate mHC to strengthen conventional residual connections, enhancing stability of signal propagation across layers while preserving model expressivity.
Muon Optimizer: We employ the Muon optimizer for faster convergence and greater training stability.

Read the full model card (1,925 words)

Configuration

Architecture: DeepseekV4ForCausalLM
Context length (tokens): 1,048,576
Layers: 43
Hidden size: 4,096
Attention heads: 64
Key/value heads: 1
Head dimension: 512
Vocabulary size: 129,280
Routed experts: 256
Experts active per token: 6
Sliding window (tokens): 128
RoPE base: 10,000
Stored precision: bfloat16
Model type: deepseek_v4
Quantization: fp8

Identity and Version

Repository: deepseek-ai/DeepSeek-V4-Flash
Publisher: DeepSeek
Task: Text generation
Modality: Text
Library: transformers
Parameters: 290.9B parameters
Languages: Not stated by the source
Revision: 60d8d70770c6776ff598c94bb586a859a38244f1
First published: 2026-04-22
Last updated: 2026-06-22

Files and Weights

73 files, 159.6 GB in total. The weights are 46 files totalling 159.6 GB in safetensors.

Weights46 files · 159.6 GB

Configuration14 files · 5.5 MB

Tokenizer2 files · 6.4 MB

Documentation4 files · 23.3 KB

Other6 files · 1.0 MB

Repository1 file · 1.6 KB

Every file

File	Type	Size	SHA-256
model-00001-of-00046.safetensors	Weights	1.1 GB	517658661390
model-00002-of-00046.safetensors	Weights	3.6 GB	f04048189d3b
model-00003-of-00046.safetensors	Weights	3.6 GB	df5f80b9b4ca
model-00004-of-00046.safetensors	Weights	3.6 GB	948250b46a6f
model-00005-of-00046.safetensors	Weights	3.6 GB	9fda158bc636
model-00006-of-00046.safetensors	Weights	3.6 GB	51a65e6d9d0c
model-00007-of-00046.safetensors	Weights	3.6 GB	2d782c46d6d2
model-00008-of-00046.safetensors	Weights	3.6 GB	b7d9d8d8932e
model-00009-of-00046.safetensors	Weights	3.6 GB	3197a42d282a
model-00010-of-00046.safetensors	Weights	3.6 GB	c9cef4200444
model-00011-of-00046.safetensors	Weights	3.6 GB	916b0b34b713
model-00012-of-00046.safetensors	Weights	3.6 GB	05739c7d91a3
model-00013-of-00046.safetensors	Weights	3.6 GB	47c5e416b60b
model-00014-of-00046.safetensors	Weights	3.6 GB	c881e2671ab4
model-00015-of-00046.safetensors	Weights	3.6 GB	cb3daae8e465
model-00016-of-00046.safetensors	Weights	3.6 GB	9eba661fba31
model-00017-of-00046.safetensors	Weights	3.6 GB	0c36cbc026c5
model-00018-of-00046.safetensors	Weights	3.6 GB	1298a07452a4
model-00019-of-00046.safetensors	Weights	3.6 GB	d3687748aff7
model-00020-of-00046.safetensors	Weights	3.6 GB	906c652f3c36
model-00021-of-00046.safetensors	Weights	3.6 GB	f270bf4d0f00
model-00022-of-00046.safetensors	Weights	3.6 GB	d02261b8f1c8
model-00023-of-00046.safetensors	Weights	3.6 GB	69fab8bfa1cd
model-00024-of-00046.safetensors	Weights	3.6 GB	baba23c06a7b
model-00025-of-00046.safetensors	Weights	3.6 GB	085b7736ebe3
model-00026-of-00046.safetensors	Weights	3.6 GB	fdde6791ab71
model-00027-of-00046.safetensors	Weights	3.6 GB	2f207b9aef9c
model-00028-of-00046.safetensors	Weights	3.6 GB	2cc519b5a03a
model-00029-of-00046.safetensors	Weights	3.6 GB	d10bf34c789f
model-00030-of-00046.safetensors	Weights	3.6 GB	0f1c471fa9d9
model-00031-of-00046.safetensors	Weights	3.6 GB	beafa59d64fa
model-00032-of-00046.safetensors	Weights	3.6 GB	5c6b2934d87a
model-00033-of-00046.safetensors	Weights	3.6 GB	c05f917e873d
model-00034-of-00046.safetensors	Weights	3.6 GB	666b77201ec6
model-00035-of-00046.safetensors	Weights	3.6 GB	5e6b9a54fd14
model-00036-of-00046.safetensors	Weights	3.6 GB	ab72ad9d171f
model-00037-of-00046.safetensors	Weights	3.6 GB	93d68bcfc36f
model-00038-of-00046.safetensors	Weights	3.6 GB	809fb799edcf
model-00039-of-00046.safetensors	Weights	3.6 GB	49dba2489174
model-00040-of-00046.safetensors	Weights	3.6 GB	09a7b8b6957f
model-00041-of-00046.safetensors	Weights	3.6 GB	a564ac6c6cc7
model-00042-of-00046.safetensors	Weights	3.6 GB	bd3f5b898b04
model-00043-of-00046.safetensors	Weights	3.6 GB	85a414c7991c
model-00044-of-00046.safetensors	Weights	3.6 GB	438b052b8a2d
model-00045-of-00046.safetensors	Weights	1.1 GB	9a0fd242134e
model-00046-of-00046.safetensors	Weights	3.6 GB	f58f722893a6
config.json	Configuration	1.7 KB	—
encoding/encoding_dsv4.py	Configuration	27.9 KB	—
encoding/test_encoding_dsv4.py	Configuration	3.7 KB	—
encoding/tests/test_input_1.json	Configuration	2.7 KB	—
encoding/tests/test_input_2.json	Configuration	526 B	—
encoding/tests/test_input_3.json	Configuration	4.5 KB	—
encoding/tests/test_input_4.json	Configuration	2.7 KB	—
generation_config.json	Configuration	170 B	—
inference/config.json	Configuration	991 B	—
inference/convert.py	Configuration	7.1 KB	—
inference/generate.py	Configuration	6.3 KB	—
inference/kernel.py	Configuration	22.2 KB	—
inference/model.py	Configuration	38.6 KB	—
model.safetensors.index.json	Configuration	5.4 MB	—
LICENSE	Documentation	1.1 KB	—
README.md	Documentation	13.1 KB	—
encoding/README.md	Documentation	8.1 KB	—
inference/README.md	Documentation	951 B	—
assets/dsv4_performance.png	Other	1.0 MB	8fd472981a4c
encoding/tests/test_output_1.txt	Other	2.4 KB	—
encoding/tests/test_output_2.txt	Other	342 B	—
encoding/tests/test_output_3.txt	Other	3.3 KB	—
encoding/tests/test_output_4.txt	Other	2.6 KB	—
inference/requirements.txt	Other	92 B	—
.gitattributes	Repository	1.6 KB	—
tokenizer.json	Tokenizer	6.4 MB	—
tokenizer_config.json	Tokenizer	801 B	—

License and Download

License: mit
Access: Open weights, no gate
Download size: 159.6 GB

Download from DeepSeek

Released by DeepSeek through its official repository on Hugging Face. Read the license.

Built From

Described by arXiv:2606.19348

Evaluations

Each result is shown as reported, with the conditions its reporter stated. None is a SAVRN measurement. A comparison lines two results up only when their configuration, unit and setup are all stated and identical.

Benchmark	Conditions	Result	Reported by	Revision	Date
Idavidrein/gpqa	Task diamondMetric diamondComparison conditions not established	88.1	Model Card Reported by a third party	Evaluated revision not stated	2026-04-24
IntelligenceLab/Long-Horizon-Terminal-Bench	Task lhtb_solvedMetric lhtb_solvedSetup 2/46 tasks solved at reward >= 0.95, estimated at the leaderboard's 90-minute budget from interim verifier checkpoints of a 3-hour terminus-2 run; NOT a real 90-minute run. Range 2-3/46 (riscv-core-debug unresolved). The same run scores 5/46 at its full 3-hour budget. Official LHTB Harbor harness.Comparison conditions not established	2	LHTB run artifacts Reported by a third party	Evaluated revision not stated	2026-08-08
SWE-bench/SWE-bench_Multilingual	Task swe_bench_multilingual_%_resolvedMetric swe_bench_multilingual_%_resolvedComparison conditions not established	73.3	Model Card Reported by a third party	Evaluated revision not stated	2026-08-10
SWE-bench/SWE-bench_Verified	Task swe_bench_%_resolvedMetric swe_bench_%_resolvedComparison conditions not established	79	Model Card Reported by a third party	Evaluated revision not stated	2026-04-24
TIGER-Lab/MMLU-Pro	Task mmlu_proMetric mmlu_proComparison conditions not established	86.4	Model Card Reported by a third party	Evaluated revision not stated	2026-04-24
benchflow/skillsbench	Task skillsbench_v1_1Metric skillsbench_v1_1Setup with-skills; BenchFlow harness; OpenHands agent; 87 tasks x 3 trials; full 261/261 coverageComparison conditions not established	44.7	SkillsBench v1.1 official leaderboard Reported by a third party	Evaluated revision not stated	2026-06-11
cais/hle	Task hleMetric hleComparison conditions not established	34.8	Model Card Reported by a third party	Evaluated revision not stated	2026-04-24
claw-eval/Claw-Eval	Task generalMetric generalSetup Pass³% \| N=3 \| 161 tasksComparison conditions not established	57.8	Claw-Eval Leaderboard Reported by a third party	Evaluated revision not stated	2026-04-23
claw-eval/Claw-Eval	Task multi_turnMetric multi_turnSetup Pass³% \| N=3 \| 38 tasksComparison conditions not established	57.9	Claw-Eval Leaderboard Reported by a third party	Evaluated revision not stated	2026-04-23
harborframework/terminal-bench-2.0	Task terminalbench_2Metric terminalbench_2Comparison conditions not established	56.9	Model Card Reported by a third party	Evaluated revision not stated	2026-04-24
joelniklaus/LEXam-hard	Task lexam_hardMetric lexam_hardSetup lighteval, LEXam paper prompts, one response per question, no tools; DeepSeek-R1-0528 judge; mean of the German and English means over the 518 questions, 0-100Comparison conditions not established	38.19	SwissLegalEvals per-sample details (lighteval) Reported by a third party	Evaluated revision not stated	2026-06-12

Memory Requirements

Precision	Weights in memory
As published	159.6 GB
16-bit	581.9 GB
8-bit	290.9 GB
4-bit	145.5 GB

Weights only, from the published parameter count; the key-value cache and runtime add to this.

Hosted Prices

Host	Input / output	Unit	Observed
DeepInfra	$0.09 / $0.18	input / output, per million tokens	Sep 18, 2026
Novita	$0.14 / $0.28	input / output, per million tokens	Sep 18, 2026
Vultr	$0.30 / $1.00	input / output, per million tokens	Sep 10, 2026

From the SAVRN Index.

Built on This Model

Quantized fromdeepseek-v4-gguf
Derived fromdeepseek-v4-gguf

Questions About DeepSeek-V4-Flash

How much GPU memory does DeepSeek-V4-Flash need?

About 698.3 GB at 16-bit and 174.6 GB at 4-bit: the weights (290.9B parameters) plus a working margin. A long context needs more.

What is the cheapest GPU to run DeepSeek-V4-Flash on?

At 16-bit, 3x MI325X from $6.00 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.

Can I use DeepSeek-V4-Flash commercially?

Yes. DeepSeek-V4-Flash is released under MIT License. The MIT License is a short permissive license. It permits commercial use, modification and redistribution, provided the copyright notice and permission notice are included.

What is DeepSeek-V4-Flash's context length?

1,048,576 tokens, from the maximum position embeddings in its published configuration.

Similar Models

Model · Text generation

DeepSeek-V4-Flash-0731

DeepSeek

DeepSeek-V4-Flash-0731 is the official release of DeepSeek-V4-Flash, superseding the preview version, with substantially enhanced agentic capabilities. It has the same model structure as DeepSeek-V4-Flash-DSpark, i.e. it comes with a speculative decoding module attached. DeepSeek-V4-Flash-0731 outperforms DeepSeek-V4-Pro (Preview) on benchmarks listed below despite its far smaller activated parameter count, and is broadly competitive with the strongest proprietary models available. 1. For the Code Agent tasks among the public benchmarks above, DeepSeek-V4-Flash-0731 is evaluated with the minimal mode of DeepSeek Harness (to be released) as the agent framework, using the max reasoning effort…

Open weights mit 304.2B parameters 1,048,576 tokens transformers

View model

Model · Text generation

MiniMax-M2.7

MiniMax

Join Our WeChat Discord community. MiniMax Agent API CLI MiniMax Website Hugging Face GitHub ModelScope LICENSE MiniMax-M2.7 is our first model deeply participating in its own evolution. M2.7 is capable of building complex agent harnesses and completing highly elaborate productivity tasks, leveraging Agent Teams, complex Skills, and dynamic tool search. For more details, see our blog post. M2.7 initiates a cycle of model self-evolution: during development, we let the model update its own memory, build dozens of complex skills for RL experiments, and improve its own learning process based on experiment results. An internal version of M2.7 autonomously optimized a programming scaffold over…

Open weights other 228.7B parameters 204,800 tokens transformers

View model

Model · Text generation

GLM-5.2-NVFP4

NVIDIA

The NVIDIA GLM-5.2 NVFP4 model is the quantized version of ZAI’s GLM-5.2 model, which is an auto-regressive language model that uses an optimized transformer architecture. GLM-5.2 is a Mixture-of-Experts (MoE) model for reasoning and coding that uses sparse attention (with an IndexShare indexer) to support a long context. For more information, please check here. The NVIDIA GLM-5.2 NVFP4 model is quantized with Model Optimizer. This model is ready for commercial or non-commercial use. GOVERNING TERMS: Use of the model is governed by the MIT License, same as the base model. Global Developers looking to take off-the-shelf, pre-quantized models for deployment in AI Agent systems, chatbots, RAG…

Open weights mit 381B parameters 1,048,576 tokens Model Optimizer

View model

Model · Text generation

DeepSeek-V4-Flash-DSpark

DeepSeek

Note: DeepSeek-V4-Flash-DSpark is not a new model. It is the same checkpoint with an additional speculative decoding module attached. A minimal inference example is available in the inference folder. For more details, refer to: https://github.com/deepseek-ai/DeepSpec We present a preview version of DeepSeek-V4 series, including two strong Mixture-of-Experts (MoE) language models — DeepSeek-V4-Pro with 1.6T parameters (49B activated) and DeepSeek-V4-Flash with 284B parameters (13B activated) — both supporting a context length of one million tokens. DeepSeek-V4 series incorporate several key upgrades in architecture and optimization: 1. Hybrid Attention Architecture: We design a hybrid…

Open weights mit 165.3B parameters 1,048,576 tokens transformers

View model

Model · Text generation

NVIDIA-Nemotron-3-Super-120B-A12B-BF16

NVIDIA

For more details on how to deploy and use the model - see the Quick Start Guide below! The post-training data has a cutoff date of February 2026. The pre-training data has a cutoff date of June 2025. NVIDIA Nemotron™ is a family of open models with open weights, training data, and recipes, delivering leading efficiency and accuracy for building specialized AI agents. Nemotron-3-Super-120B-A12B-BF16 is a large language model (LLM) trained by NVIDIA, designed to deliver strong agentic, reasoning, and conversational capabilities. It is optimized for collaborative agents and high-volume workloads such as IT ticket automation. Like other models in the family, it responds to user queries and…

Open weights other 123.6B parameters 262,144 tokens transformers

View model

Model · Text generation

gpt-oss-120b

OpenAI

Welcome to the gpt-oss series, OpenAI’s open-weight models designed for powerful reasoning, agentic tasks, and versatile developer use cases. We’re releasing two flavors of these open models: - gpt-oss-120b — for production, general purpose, high reasoning use cases that fit into a single 80GB GPU (like NVIDIA H100 or AMD MI300X) (117B parameters with 5.1B active parameters) - gpt-oss-20b — for lower latency, and local or specialized use cases (21B parameters with 3.6B active parameters) Both models were trained on our harmony response format and should only be used with the harmony format as it will not work correctly otherwise. You can use gpt-oss-120b and gpt-oss-20b with Transformers.…

Open weights apache-2.0 116.8B parameters 131,072 tokens transformers

View model