DeepSeek V4 Flash

About

This is the official V4 Flash release, which replaces the earlier preview. DeepSeek combined two compressed ways of reading text to make long inputs more efficient.

How it compares

Of the 7 chat models Outmake recommends for a 192 GB Mac, Outmake estimates it is 4th fastest. Gemma 4 26B is about 1.5 times as fast.

Five of those 7 read pictures, including Qwen3.8 Flash Next, which Outmake estimates writes at about the same speed.

In LM Arena's public vote it ranks 12th of 59 open models for coding and 13th of 62 for text.

Good to know

It was trained on more than 32 trillion pieces of text.

DeepSeek trained it with a method called Muon, chosen for faster and more stable training.

Sources: deepseek-ai/DeepSeek-V4-Flash-0731 model card, DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence, Outmake's speed estimates, LM Arena and Hugging Face. Every line comes from one of these.

Run It Anywhere

DeepSeek V4 Flash is an open model, so any app that runs GGUF files can run it. These take the UD-Q4_K_XL size from unsloth straight from Hugging Face.

• LM StudioMac, Windows, LinuxOpen
• JanMac, Windows, LinuxOpen
• MstyMac, Windows, LinuxOpen
• OllamaMac, Windows, Linuxollama run hf.co/unsloth/DeepSeek-V4-Flash-0731-GGUF:UD-Q4_K_XLCopy
• llama.cppMac, Windows, Linuxllama-server -hf unsloth/DeepSeek-V4-Flash-0731-GGUF:UD-Q4_K_XLCopy
• Docker Model RunnerMac, Windows, Linuxdocker model run hf.co/unsloth/DeepSeek-V4-Flash-0731-GGUF:UD-Q4_K_XLCopy