AI
DeepSeek V4 Flash: Building a Spider-Man Cafe Tracker
August 20267 min read
Cafe Tracker is a live, real-time map of coffee shops and eateries around BSD City and Gading Serpong in Tangerang Selatan, Indonesia. Its interface is inspired by the Spidey Tracker app from Spider-Man: Brand New Day, and every place links to Google Maps for directions and hours. You can try it at cafe-tracker.matthewswong.com.
Almost all of the app was built with DeepSeek V4 Flash 0731, the refined 0731 release of DeepSeek's open-weight V4-Flash model. It handled the heads-up display, the map and marker logic, the sector filter, and the Google Maps wiring across the whole project.
DeepSeek released the V4 series, including V4-Flash, under the permissive MIT License with open weights published on Hugging Face. That means you can download, run, and fine-tune the model yourself if you have the hardware, or use DeepSeek's hosted API.
It refers to taking an existing model and refining it again for better behavior, rather than training a brand-new model from scratch. DeepSeek's 0731 build is the V4-Flash model re-tuned for stronger agentic coding, and because the weights are open, others can fine-tune it further for their own use cases.
It is live at cafe-tracker.matthewswong.com. Open it on a phone or desktop, filter by the BSD City or Gading Serpong sector, and tap any sighting to jump to its Google Maps listing.
Key Takeaway
Cafe Tracker is a live map of coffee shops and eateries around BSD City and Gading Serpong, built by Matthews Wong with DeepSeek V4 Flash 0731. Its heads-up display is inspired by the Spidey Tracker app from the film Spider-Man: Brand New Day. The project shows how a refined open-weight model can ship a real, polished app fast.
Spider-Man: Brand New Day landed on July 31, 2026, and the detail that stuck with me was not a fight scene. It was Ned Leeds building an app he calls the Spidey Tracker to pin down where Spider-Man is. I wanted that same neon, real-time hunt interface, but pointed at something I actually chase every week: good coffee near me.
So I built Cafe Tracker, a real-time sightings map for cafes and eateries around BSD City and Gading Serpong, and I built almost all of it with DeepSeek V4 Flash 0731. The result genuinely surprised me. This post is what I made, how the model held up, and why V4 Flash is a nice example of iterative re-fine-tuning paying off.
See it live: cafe-tracker.matthewswong.com
In the film, the Spidey Tracker is a glowing dashboard of live sightings on a city map. I kept the visual language and swapped the target. Instead of one masked hero, Cafe Tracker plots dozens of cafe and eatery sightings across my two home sectors in Tangerang Selatan: BSD City and Gading Serpong.
The interface is a deliberate homage: a dark heads-up display, animated markers, a live sightings feed, and a sector filter. Every marker links straight to Google Maps for directions, hours, and the menu, so the theme never gets in the way of actually finding a table.
DeepSeek is a Chinese AI lab, backed by the hedge fund High-Flyer, that has become known for releasing capable open-weight models. In April 2026 it previewed its V4 series, including DeepSeek-V4-Flash at 284 billion parameters and the larger V4-Pro, both under the permissive MIT License and both with a one-million-token context window.
The 0731 build is the refined release of V4 Flash, and DeepSeek highlights stronger agentic coding over the preview it replaced. Open weights, a huge context window, and that agentic focus are exactly what I wanted for a from-scratch build, so I handed it the whole app and let it work.
I described the Spidey Tracker vibe, the two sectors, and the data I had, then iterated. What stood out was how well the model held the whole app in its head, from the map and the marker animation to the sector filter and the Google Maps wiring, without me re-explaining the design every few messages.
Here is my honest read. DeepSeek did not stop at the V4 preview; the 0731 build is the same Flash model taken back and refined for stronger agentic behavior. That loop of ship, learn, re-tune, and ship again is exactly what made the model feel so capable in a hands-on coding task.
Because the weights are MIT-licensed and open, that refinement is not locked away either. Anyone with the hardware can download V4 Flash and fine-tune it further for their own domain. To me, V4 Flash is a clean example of how re-fine-tuning an open base, rather than chasing one giant frozen model, compounds into something you can actually build on.
One project is not a benchmark. This is my subjective experience on a single app, and model quality varies a lot by task. Open-weight output still needs review and testing, and DeepSeek is a Chinese lab, so if you use the hosted API rather than self-hosting the open weights, factor in where your data goes.
Cafe Tracker is live now, tracking sightings across BSD City and Gading Serpong. If you are in the area, open it, hunt down your next coffee, and let me know which sector or feature I should add next.
The Spidey Tracker was the excuse, but DeepSeek V4 Flash 0731 was the real surprise. A refined, open-weight model turned a movie-night idea into a working app faster than I expected, and that is a strong argument for the re-fine-tuning approach behind it.