Run DeepSeek V4.1 Flash on a Mac.

1How much memory it needs

The model has 552 billion parameters, but about 189 GiB of its download is two Engram lookup tables that are read a few rows per token. Runtimes that leave those tables on the SSD need far less memory than the download size; builds that load everything need it all (antirez on Hugging Face).

Leave room for macOS and for the context: at about 890 bytes per token, a 100,000-token conversation adds under 0.1 GB, so the weights are what decide.

2Which Mac can run it

Keep the model on the Mac's internal SSD: Engram rows and streamed weights are read from disk during generation, and an external drive is much slower. Plan for the download too: 250 to 520 GB, plus free space while files are joined.

3How fast it runs

The first public benchmark on a Mac ran a mixed 4- and 8-bit MLX build of the September weights on an M3 Ultra Mac Studio with 512 GB (benchmark on GitHub):

That is far below what the hardware allows. Writing reads about 16 billion parameters per token, about 9 GB at 4 bits, so memory bandwidth caps one chat at roughly 90 tokens a second on an M3 Ultra (819 GB/s) and 130 on an M5 Ultra (1.2 TB/s). Support for the new architecture is weeks old, so expect newer runtimes to close part of the gap; compare current numbers on MLX Bench and in the runtimes' release notes.

Reading prompts is limited by compute rather than memory. M5-generation chips have neural accelerators in the GPU and read prompts several times faster than earlier chips with MLX (Apple).

4How to run it

DwarfStar (ds4): the leanest on memory

DwarfStar is an open-source engine built for DeepSeek Flash models, with Metal support for Macs from 96 GB and SSD streaming for smaller ones. It keeps the Engram tables on disk in every mode (GitHub).

git clone https://github.com/antirez/ds4.git
cd ds4 && make
./download_model.sh ds41f-q2
./ds4 -m gguf/DeepSeek-V4.1-Flash-Q2.gguf --ssd-streaming --ctx 32768

Leave out --ssd-streaming when the Mac has 256 GB or more. ./ds4-server serves the model over a local API.

llama.cpp: the 1017 GGUF files

Mainline llama.cpp and Ollama don't support V4.1 Flash yet. The 1017 GGUF files were made with a llama.cpp branch, which you build yourself; keeping the Engram tables on the CPU side is recommended (Hugging Face). The files are text-only for now.

git clone https://github.com/vcruz305/llama.cpp && cd llama.cpp
git checkout 2076412
cmake -B build && cmake --build build -j
./build/bin/llama-server -hf mrfakename/DeepSeek-V4.1-Flash-1017-GGUF:Q2_K -ot "engram_embd.weight=CPU"

MLX: the 1017 builds for Apple Silicon

The MLX community builds of 1017 keep DeepSeek's original FP4 and FP8 precision for most weights and include the multi-token prediction weights. They load through mlx-vlm, which also handles images, and LM Studio lists them too. They need a 512 GB Mac.

pip install -U mlx-vlm
python -m mlx_vlm.generate --model mlx-community/DeepSeek-V4.1-Flash-1017-oQ4e-mtp --prompt "Hello" --max-tokens 200

New to local models? Start with running an LLM on a Mac and MLX vs Ollama vs llama.cpp.

5No 512 GB Mac? Your options

• Use the API. DeepSeek charges $0.15 per million input tokens and $0.60 per million output tokens for deepseek-flash; see DeepSeek V4.1 Flash 1017.
• Run a smaller open model. A 64 GB Mac runs 70B-class models and a 128 GB Mac runs gpt-oss-120b well: see the best Mac for local LLMs.
• Join Macs together. Two or more Macs over Thunderbolt 5 can hold models no single one can: see Mac AI clusters.
• Let your Mac earn. MLX Node runs AI jobs for MLX Network on any Apple Silicon Mac and earns MLX Tokens, up to 1 MLX a day without a dedicated seat. Mined MLX is locked for 30 days, less with a dedicated seat; see the risks.

6Questions

How much RAM do I need for DeepSeek V4.1 Flash?

About 163 GB for the 2-bit DwarfStar build with the Engram tables left on the SSD, about 316 GB for its 4-bit build, and the whole download, 343 to 432 GB, for the MLX builds. Add room for macOS and the context.

Can a MacBook Pro run DeepSeek V4.1 Flash?

Only a 128 GB one, and only by streaming the 2-bit build from the SSD with DwarfStar, which is experimental and slow. It runs fully in memory on a Mac Studio with 256 or 512 GB.

How fast is DeepSeek V4.1 Flash on an M3 Ultra?

The first public run on a 512 GB M3 Ultra wrote about 6.5 tokens a second and read prompts at about 275 tokens a second, with an MLX build. The memory bandwidth would allow about 90 tokens a second, so newer software should be faster.

Is there an MLX version of DeepSeek V4.1 Flash 1017?

Yes. The MLX community published 1017 builds of 343 GB and 432 GB. They run with mlx-vlm or LM Studio on a Mac with 512 GB of memory.

Does DeepSeek V4.1 Flash work in Ollama?

Not in standard Ollama or llama.cpp yet. The GGUF files need a llama.cpp branch built from source, or you can use DwarfStar, which supports the model on Macs.