Hotshot just launched a 10-second AI video generator built by 4 engineers

Hotshot, a four-person AI startup, just launched an early preview of its ambitious new text-to-video generator.

Anyone can try the model for free right now at hotshot.co. It produces clips up to 10 seconds long at 720p resolution without watermarks, though free users are currently capped at two generations per day.

Four engineers versus thousands of failing H100s

Founded by Aakash Sastry, John Mullan, and Duncan Crawbuck, alongside engineer Chaitu Aluru, Hotshot spent 11 years building consumer apps before pivoting to AI video. The company previously released Hotshot-XL, an open-source model generating 1-second clips at 8 fps with over 20,000 monthly users, followed by Hotshot Act-One for 3-second clips.

The new self-titled flagship is on a whole different level. The team trained the model across four months using 600 million video clips and thousands of GPUs, backed by investors including Alexis Ohanian, Lachy Groom, and SV Angel.

Running thousands of Nvidia H100s at full throttle turned out to be an operational nightmare. According to the team's research paper, GPUs failed constantly under maximum load, requiring 24/7 human babysitting that consumed one engineer's life entirely for a month straight. To pull it off, they engineered a novel autoencoder that compresses video both spatially and temporally, drastically slashing file sizes while preserving core dynamic data.

In testing, the output shows promise across artistic styles like comic book aesthetics and rotoscoping, though prompt adherence can be hit or miss. A test prompt for a unicorn riding through Paris generated a clean, realistic clip of a horse in Paris, entirely forgetting the horn. It also noticeably lags behind heavyweights like Runway Gen-3 Alpha Turbo in overall crispness.

Why it matters

For creators and casual users, Hotshot delivers immediate utility: 10-second generations completely free of watermarks. Two daily runs are limited, but it is great for quick visual prototyping, B-roll concepts, or memes without pulling out a credit card.

For the broader AI ecosystem, Hotshot punches way above its weight. While giants pour hundreds of researchers and tens of millions into video models, four engineers managed to train a competent 10-second generator in four months. It proves clever data compression and efficient pipelines can still challenge brute-force scaling.

That said, Hotshot faces a tough uphill climb. Prompt comprehension issues like mistaking unicorns for horses show the model still lacks deep semantic precision. With Runway, Luma, and soon Sora aggressively raising the bar on resolution, camera motion, and audio integration, Hotshot needs to upgrade its fidelity quickly before its David-versus-Goliath momentum fades.