Gangsta AI
DeepSeek Just Gave the World a Free Set of Eyes — and Barely Said a Word
Open Weights

By Beck · 2026-09-01 · 4 min read

There was no keynote. No countdown, no smug demo reel, no CEO in a black tee promising to change your life. On August 31, DeepSeek just… uploaded a folder to Hugging Face and wandered off.
The folder was DeepSeek-V4-Flash-Vision-Exp — a roughly 305-billion-parameter experimental model, the company's first native vision system, bolting an image encoder onto its V4-Flash base so the thing can finally *see*. It reads charts, screenshots, photos, diagrams. And it's under an MIT license, which in plain English means: take it, run it, sell it, do whatever, don't call us.
“A frontier lab shipped a model that can see, priced it at zero, and didn't even throw a party.”
That's the part that keeps happening. Every few weeks a Chinese lab — DeepSeek, Alibaba's Qwen, Z.ai — quietly open-sources something that would've been a $200-million product launch in San Francisco, and the response back home is a round of nervous group chats. DeepSeek already torched the pricing floor on coding models earlier this summer. Now it's doing it to vision, the last thing the big closed labs were still charging a premium for.
The catch nobody mentions
Here's the thing the hype cycle skips: "experimental" is doing real work in that name. It's a first checkpoint, a base plus a freshly-glued vision aligner — the kind of model that dazzles on one image and hallucinates a third arm on the next. Open weights don't mean *good*, they mean *available*. Whether it actually beats Gemini or GPT-5.2 at reading your messy whiteboard photo is an open question, and the leaderboards will be arguing about it for a month.
Which is the whole problem with living through this, honestly. The "best" vision model changes faster than you can renew a subscription. One week it's a closed model from a trillion-dollar company, the next it's a free download from Hangzhou. Pick a favorite and you're basically betting on a horse mid-race, blindfolded.
So don't pick
The move isn't to crown a winner every Tuesday. It's to stop crowning winners. Hand the same image — the same question — to ChatGPT, Claude, Gemini, Grok, a fresh DeepSeek drop, and 30 more, and let them fight it out in front of you. Where they agree, you can relax. Where they split, *that's* the interesting part, and it's exactly the thing a single model would've hidden from you with a confident wrong answer. Gangsta AI runs that gauntlet for you and fuses it into one cross-checked, cited verdict — so a free model release in China is just another voice in the room, not a reason to switch everything.
DeepSeek gave everybody a free set of eyes this weekend. Might as well use more than one.
Curious which model actually leads for the thing *you* do — vision, coding, writing, research? Start with our running best AI models breakdown and let the field argue before you commit.
Sources / Receipts
More: Best AI models · Compare all AI · Frontier Models · All articles