DeepSeek-V4-Flash

Updated
24.08.2026
Thinking
Reasoning
Code
Multilingual

DeepSeek-V4-Flash is a 284B MoE model with 13B active parameters and a 1M token context, released by DeepSeek under the MIT license.

At a glance

  • License: MIT
  • Parameters: 284B total, 13B active
  • Context length: 1M tokens
  • Modalities: Text
  • Minimum hardware: 96 GB of RAM or unified memory (smallest GGUF is 82.6 GB)

What is DeepSeek-V4-Flash?

DeepSeek-V4-Flash is a 284B parameter Mixture-of-Experts model from DeepSeek, published as a preview of the V4 series alongside the 1.6T DeepSeek-V4-Pro. It activates 13B parameters per token and supports a context of one million tokens. It is the smaller of the two V4 models and the one within reach of a big workstation: the smallest GGUF build is 82.6 GB. The weights went up on Hugging Face on April 22, 2026 under the MIT license.

SpecificationDeepSeek-V4-Flash
Total parameters284B
Activated parameters13B per token
ArchitectureMoE, hybrid attention (CSA + HCA)
Context window1M tokens
ModalitiesText
Reasoning modesNon-think, Think High, Think Max
Pre-training dataMore than 32T tokens
PrecisionFP4 + FP8 mixed (MoE experts in FP4, most other weights FP8)
Release dateApril 22, 2026
LicenseMIT

The million-token context comes from a hybrid attention design that pairs Compressed Sparse Attention with Heavily Compressed Attention. DeepSeek quotes the larger V4-Pro at 27% of the single-token inference FLOPs and 10% of the KV cache of DeepSeek-V3.2 at a 1M-token context, and both V4 models share the architecture. Training used the Muon optimizer, Manifold-Constrained Hyper-Connections to strengthen conventional residual connections, and more than 32T tokens of pre-training data. Post-training ran in two stages: domain experts trained separately with SFT and RL, then merged into a single model through on-policy distillation. One practical note: the repo ships Python encoding scripts instead of a Jinja chat template, which matters if you wire the model up yourself but not if an app handles the template for you.

DeepSeek-V4-Flash benchmarks

DeepSeek published launch numbers on the model card comparing V4-Flash across its three reasoning modes, with the 1.6T V4-Pro in Think Max mode as the reference:

BenchmarkV4-Flash Non-ThinkV4-Flash HighV4-Flash MaxV4-Pro Max
MMLU-Pro
Academic knowledge
83.086.486.287.5
GPQA Diamond
Expert science
71.287.488.190.1
LiveCodeBench
Competitive coding
55.288.491.693.5
HMMT Feb 2026
Olympiad math
40.891.994.895.2
MRCR 1M
Long context
37.576.978.783.5
SWE-bench Verified
Software engineering
73.778.679.080.6
Terminal Bench 2.0
Terminal agents
49.156.656.967.9

Think Max brings the 13B-active Flash within a point or two of the flagship on math and code: 94.8 against 95.2 on HMMT, 91.6 against 93.5 on LiveCodeBench. The gap is wider on agentic work, where Pro Max leads Terminal Bench 2.0 by 11 points. The Non-think column shows how much of the score comes from thinking: 40.8 to 94.8 on HMMT.

DeepSeek-V4-Flash hardware requirements

The system requirement to check is memory. DeepSeek ships the original weights in mixed FP4 + FP8; the GGUF builds below come from unsloth/DeepSeek-V4-Flash-GGUF, each split into shards of up to 50 GB, listed here as totals.

MemoryBuild to pickFile size
96 GBUD-IQ1_S82.6 GB
128 GBUD-IQ3_XXS103.0 GB
192 GBUD-Q4_K_XL155.1 GB
256 GB and upUD-Q8_K_XL161.9 GB

When two builds both fit, take the larger one. Only 13B of the 284B parameters are active per token, so generation speed holds up better than the file sizes suggest once the weights are in memory. Budget headroom for context on top of the weights: DeepSeek recommends at least 384K tokens of context for Think Max mode. If the format is new to you, start with what GGUF is.

How to run DeepSeek-V4-Flash in Atomic Chat

Atomic Chat is a free local app for macOS, Windows and Linux. It includes a Hugging Face model browser and a built-in chat, with no manual llama.cpp build required.

  1. Download Atomic Chat for your platform and open it.
  2. Search for DeepSeek-V4-Flash in the model browser and open Download Options.
  3. Pick the build that fits the memory you have, then start a chat.

For the rest of the lineup, see every DeepSeek model you can run locally, or the 1.6T flagship DeepSeek-V4-Pro.

DeepSeek-V4-Flash license

DeepSeek-V4-Flash is released under the MIT license, one of the most permissive licenses in open source. It permits commercial use, modification and redistribution with no royalties, so you can build products on top of the model and run it on your own hardware without a usage fee.

Get the weights from Hugging Face

huggingface-cli download deepseek-ai/DeepSeek-V4-Flash
from transformers import AutoModel
model = AutoModel.from_pretrained("deepseek-ai/DeepSeek-V4-Flash")
Desktop
macOS
(Intel and Apple Silicon)
Download
Windows
(x64)
Download
Windows
(ARM64)
Download
Linux
(x86_64)
Download

Frequently asked questions

DeepSeek-V4-Flash is an open-weight Mixture-of-Experts language model from deepseek-ai, part of the deepseek_v4 family. It has a 158.1B parameter count, ships in FP8, and supports a 1,048,576-token context window. Its listed capabilities are thinking, reasoning, code, and multilingual, and it is built so only some experts activate per token to keep inference efficient.

The full FP8 weights are 158.1B parameters, so you need a machine with substantial memory, typically a high-VRAM GPU setup or a workstation with large unified memory. The MoE design keeps active compute per token lower than the headline size implies, which helps. Atomic Chat handles the download and runtime, so you mainly need enough RAM and VRAM to hold the model and your context.

Yes. The weights are released under the MIT license, and running the model locally through Atomic Chat costs nothing per token. There is no API key or subscription. You only pay for the hardware and electricity to run it on your own machine.

Yes. Once you download the weights with Atomic Chat, the model runs fully on your device with no internet connection required. Prompts and responses stay local, which keeps your data private. You only need a connection for the initial download of the model files.

You can download the weights from the deepseek-ai repository on Hugging Face with huggingface-cli download deepseek-ai/DeepSeek-V4-Flash, then serve them using Transformers or vLLM. The simpler path is to open Atomic Chat and load DeepSeek-V4-Flash with one click, which sets up the download and runtime for you. After that you chat with it directly, offline and on your own hardware.