Antirez Brings Vision to DS4, DeepSeek V4 Flash Runs Locally on an M5 Max

wpnews.pro

antirez showed DeepSeek V4 Flash with vision running fast and fully local on a Mac M5 Max, and finished the Metal, CUDA and ROCm backends of his DS4 engine.

[Most Read](https://pasqualepillitteri.it/en/news/576/claude-code-skills-design-uiux-guide)

[20 AI News & Trends](https://pasqualepillitteri.it/en?cat=77)

[1 Automotive Tech](https://pasqualepillitteri.it/en?cat=386)

[1 Intelligenza Artificiale](https://pasqualepillitteri.it/en?cat=22)

[13 Google AI & Gemini](https://pasqualepillitteri.it/en?cat=67)

[2 Guides & Tutorials](https://pasqualepillitteri.it/en?cat=18)

[1 AI for Professionals](https://pasqualepillitteri.it/en?cat=82)

[19 Cybersecurity](https://pasqualepillitteri.it/en?cat=21)

[1 Apple](https://pasqualepillitteri.it/en?cat=157)

[2 Reports & Analysis](https://pasqualepillitteri.it/en?cat=57)

[6 Claude Code & Anthropic](https://pasqualepillitteri.it/en?cat=62)

[1 Finance](https://pasqualepillitteri.it/en?cat=87)

[1 Agent Infrastructure](https://pasqualepillitteri.it/en?cat=108)

[1 Diritto & Normative](https://pasqualepillitteri.it/en?cat=142)

[5 OpenAI & ChatGPT](https://pasqualepillitteri.it/en?cat=147)

[3 Videogiochi](https://pasqualepillitteri.it/en?cat=414)

[2 Software Development](https://pasqualepillitteri.it/en?cat=23)

[1 Opinions](https://pasqualepillitteri.it/en?cat=127)

1.What antirez showed
2.What DS4 is, and why one engine just for DeepSeek V4 Flash
3.Metal, CUDA and ROCm, the same vision on Mac, NVIDIA and AMD
4.The numbers on M5 Max
5.What it takes to run it at home
6.Key Takeaways
7.Frequently Asked Questions (FAQ)
8.The bottom line
9.Rate this article
10.Related Articles
11.Looking for a Software Engineer?

Salvatore Sanfilippo, the Sicilian programmer who created Redis and whom the internet knows as antirez, on September 1, 2026 showed off DeepSeek V4 Flash with vision running fast on a Mac M5 Max and analyzing an image, then announced that Metal, CUDA and ROCm support for his DS4 engine was finished and would ship within hours. The story here isn't a new model. It's that one of the most capable multimodal models around now runs locally, on your own machine, without ever sending the image to a server.

DeepSeek v4 Flash experimental with vision running (fast) on an m5 max computer and analyzing an image. The implementation for Metal / CUDA / ROCm is finished, just doing the last tests today before releasing it.

— Salvatore Sanfilippo (@antirez)[September 1, 2026]

What antirez showed #

The video attached to the post does one thing, and does it well. A Mac M5 Max takes in an image, passes it to DeepSeek V4 Flash, and returns a description within seconds, with the model crunching everything on the Apple chip and no cloud connection involved. antirez wrote that the implementation for the three backends (Metal for Macs, CUDA for NVIDIA cards, ROCm for AMD ones) was finished, with only the last tests left before release.

What actually raises the bar isn't speed by itself, though it counts. Vision, the part almost everyone shipped off to a remote API until yesterday, now lives inside an open source engine built from a few thousand lines of C. Anyone working with confidential documents, medical scans or material under legal restrictions knows exactly how much that difference weighs. The image never leaves the machine.

What DS4 is, and why one engine just for DeepSeek V4 Flash #

DS4, short for DwarfStar4, is a specialized inference engine antirez published on GitHub on May 6, 2026, and which today counts over 21,900 stars (21,984 to be exact as of September 1) under an MIT license. It doesn't try to run just any model. It runs one model well, DeepSeek V4 Flash, with support for the PRO variant too on machines with a lot of memory. It's a sharp choice, almost swimming against the current compared to generalist runtimes like llama.cpp that aim to cover everything.

The logic behind that specialization holds up. DeepSeek V4 Flash is a mixture-of-experts model, meaning only a small fraction of its total parameters light up for each token. For Flash that means 13 billion active parameters out of 284 billion overall, an extreme sparsity. Squeezing quality and speed out of an architecture this peculiar pays off far more with kernels hand written for that exact shape than with an engine built to be good at everything. We already covered the underlying bet when DS4's roadmap and its early strategic implications for local inference came out.

The model DS4 gives vision to carries DeepSeek's signature. The Chinese lab released DeepSeek V4 Flash Vision-Exp on August 21, 2026, an experimental version that adds image understanding without losing anything on text. Images get counted as tokens, capped at 384 tokens per image, and DeepSeek claims multimodal performance close to Claude Opus 4.8. Until yesterday you used that model through an API. As of today you can keep it at home.

Metal, CUDA and ROCm, the same vision on Mac, NVIDIA and AMD #

This is the actual news. Adding vision to a single backend would have been an exercise; bringing it to all three within days is a different animal. The GitHub commits tell the story of the final sprint, with Vision-Exp support landing first on Metal, then two consecutive patches on August 31 fixing the CUDA build and the ROCm build, and finally, on September 1, usage documentation and a hardening pass on the multimodal cache.

In practice, the same image on the same model gets analyzed on the Apple laptop in the living room, on the NVIDIA workstation at the office, or on an AMD mini PC like the Framework Desktop with Strix Halo, with no three separate codebases and no three different quality tiers. For local vision, until now a patchwork of partial solutions tied to one specific GPU, a single engine spanning three accelerators changes the game.

Vision, in itself, isn't free in memory terms. The Vision-Exp model carries a separate visual encoder weighing 0.9 GiB, on top of the language model's own weights. Nothing prohibitive, but a detail worth factoring in when sizing the machine.

The numbers on M5 Max #

antirez published benchmarks on an M5 Max with 128 GB of RAM, running the Flash model at 2 bit. These are numbers worth reading with the right mindset. This isn't a data center. It's a laptop analyzing images with a model carrying almost 300 billion parameters.

| Stage | Short context (2,048 tokens) | Long context (65,536 tokens) |
|---|---|---|
Prefill (reading the prompt) |

Nearly 40 tokens per second in generation with a short context means a reply that scrolls faster than you can read it. Even at full context, with 65,000 tokens loaded, the model stays above 27 tokens per second, a pace entirely usable for real work. We'd already looked at the comparison between a local Mac and a $200-a-month cloud subscription, and the math increasingly favors the machine at home.

What it takes to run it at home #

The 2-bit quantized version of Flash takes up roughly 81 GB, so the comfortable range starts at 96 GB of unified RAM and sits well at 128 GB. DS4 only quantizes the MoE's routed experts, leaving the rest of the model intact, and that's why quality holds up even at 2 bit. Anyone without that much memory can rely on streaming weights from SSD, which lets you run models bigger than available RAM at some cost in speed.

Using vision from the command line is about as direct as it gets. In the interactive CLI, the /read image.png command alone feeds the model a PNG or JPEG file, and from there you ask questions about the image the way you would with any multimodal chat. Under the hood DS4 offers a lot more, from a persistent on-disk KV cache to tool calling, all the way to a server with an OpenAI and Anthropic compatible API and a built-in coding agent, all of it built to work with agents or from the CLI.

Worth noting, antirez builds DS4 with heavy help from models, crediting GPT 5.5, 5.6 and Claude Fable in the README, while he stays in charge of design and testing himself. This isn't the first time he's brought sharp ideas to local AI, like when he introduced Redis Array to handle enormous knowledge bases for agents.

Key Takeaways #

Vision from a near-frontier model now stays on your machine. The image DeepSeek V4 Flash analyzes never touches a server, and for anyone handling sensitive data that matters more than any benchmark.
The same model at the same quality runs on Mac, NVIDIA and AMD. antirez closed out the three backends, Metal, CUDA and ROCm, together, so your hardware choice stops dictating what you can do.
A laptop is enough to work with it seriously. On M5 Max the model generates nearly 40 tokens per second with a short context and stays above 27 even at full context, a pace that holds up for daily work.
The entry price is memory, not an exotic GPU. You need roughly 81 GB for the 2-bit model, comfortable at 96 to 128 GB, otherwise you scale onto SSD and accept less speed.
Specialization beats generality on this architecture. An engine tailored to a single sparse MoE outperforms runtimes built to be good at everything, and that's the reason DS4 exists.

Frequently Asked Questions (FAQ) #

1. Is DS4 free, and can you use it right away?

Yes, DS4 is free and open source under an MIT license. The code lives on GitHub and anyone can download it, build it and use it, commercial projects included. Vision support is the one announced on September 1, 2026, so it's worth starting from the latest version of the main branch.

2. What's the difference between DeepSeek V4 Flash and the Vision-Exp version?

Vision-Exp adds image understanding while staying identical on text. DeepSeek released it on August 21, 2026 as an experimental version. On text tasks, reasoning and agent work included, it performs like the regular Flash, while additionally reading images, counted as tokens, up to 384 per image.

3. Do you need a Mac M5 Max?

No, the M5 Max is just the demo machine. DS4 runs on Mac with Apple Silicon via Metal, on NVIDIA cards via CUDA, DGX Spark included, and on AMD systems via ROCm like the Framework Desktop with Strix Halo. Available memory matters more than the chip brand.

4. How much RAM do you really need?

Roughly 81 GB for the Flash model at 2 bit, so the ideal range sits between 96 and 128 GB. With less memory you can still use SSD weight streaming, which runs the model at the cost of speed. Vision adds a 0.9 GiB encoder on top of the memory requirement.

5. Why not use a generic engine like llama.cpp?

Because DeepSeek V4 Flash is a sparse MoE and benefits from custom-built kernels. antirez chose to hand optimize for that single architecture instead of covering every model, and on Flash that translates into more speed and more quality than a universal runtime delivers.

The bottom line #

For years local AI chased the cloud and stayed a step behind, fine for tinkering, not for real work. This release moves the boundary another notch. Not because antirez invented something that didn't exist before, but because he made ordinary something that was hobbyist territory only yesterday. A near-frontier multimodal model that sees images, on your own machine, identical across three kinds of hardware, with a two-word command. It remains to be seen how well the experimental vision variant holds up and how fast stable support arrives. But the direction is clear, and it's worth keeping an eye on. If you have questions or want to share your own experience with local inference, use the contact form below.

Enjoyed the article?

Buy me a coffee

Stay updated

🌐 3.4k readers worldwide, every SaturdaySubscribe to the newsletter to receive new articles directly in your inbox.

source & further reading

pasqualepillitteri.it — original article The evolution of AI-assisted software engineering paradigms

Antirez Brings Vision to DS4, DeepSeek V4 Flash Runs Locally on an M5 Max

antirez showed DeepSeek V4 Flash with vision running fast and fully local on a Mac M5 Max, and finished the Metal, CUDA and ROCm backends of his DS4 engine.

What antirez showed #

What DS4 is, and why one engine just for DeepSeek V4 Flash #

Metal, CUDA and ROCm, the same vision on Mac, NVIDIA and AMD #

The numbers on M5 Max #

What it takes to run it at home #

Key Takeaways #

Frequently Asked Questions (FAQ) #

1. Is DS4 free, and can you use it right away?

2. What's the difference between DeepSeek V4 Flash and the Vision-Exp version?

3. Do you need a Mac M5 Max?

4. How much RAM do you really need?

5. Why not use a generic engine like llama.cpp?

The bottom line #

Related Articles

DwarfStar4 (DS4) Roadmap by antirez: DeepSeek V4 Flash on Apple Silicon and CUDA

Local AI Inference as a Strategic Choice: Why antirez's DS4 Changes the Rules in 2026

Apple launches the 512GB Mac Studio M5 Ultra, antirez explains why the missing 1TB tier is no loss for local AI

OpenAI and Anthropic Snap Up Mac Minis as Apple Posts Record Mac Quarter

Can a local Mac Studio replace your $200 AI subscription?

DeepSeek Launches V4-Flash-Vision-Exp: Cheap Model Now Sees Images, Takes On Opus 4.8

Stay updated

Run your AI side-project on zahid.host