![What People Are Building With Kimi K3 [2026]](/_next/image?url=%2Fimages%2Fai-builders%2Fkimi-k3-hero.jpg&w=3840&q=75)
Kimi K3, Moonshot's 2.8 trillion parameter open-weight model, shipped July 16 and priced itself at roughly a third of Claude Fable 5. Since then people have used it to rebuild Google Maps 3D mode in two prompts, generate a game-ready Blender fighter jet for under a dollar, and run it head to head against Fable 5 and GPT-5.6 Sol. A security researcher also reports it found real Redis zero-days and wrote a working exploit, a claim we treat carefully below. Every build links to the builder's original post. For the full model breakdown, see our Kimi K3 review.
A new model is easiest to judge by what people rush to build with it in the first week. K3 launched cheap and open, and the demos came fast. We watched each one, linked every original post, and where a claim needed checking we checked it.
Google Maps 3D, rebuilt in two prompts
Kirill rebuilt the 3D map mode in about 90 minutes.
Kirill posted an interactive 3D map of New York City in full building detail, with the whole globe zoomable on top of it. His description of the effort: "two prompts. that's the whole build."
The buildings use real height data for extrusion, and the shadows move with the sun. He also shipped a written A-to-Z guide to building with K3 alongside it.
A real Redis zero-day, exploit included
Redis confirmed the bugs. The AI-found-it story is one researcher's account.
This is the loudest K3 claim of the week, so treat it in two parts. The confirmed part: Redis shipped seven security releases on July 23 for two memory corruption flaws, and a public proof-of-concept repo now carries working exploits for several Redis versions.
The unconfirmed part: security researcher Chaofan Shou says Kimi K3 found the bugs and wrote the exploit itself, coordinating 32 agents in 27 minutes. That timing and degree of autonomy are his own account. As The Hacker News put it, the counts and the claimed autonomy "remain self-reported," and neither Redis nor Moonshot has verified them.
Not sure which AI model to use?
12 models · Personalized picks · 60 seconds
A game-ready fighter jet for under a dollar
Soya wired K3 to Blender through an MCP tool.
Soya gave K3 one instruction: use the Blender MCP tool to make a fighter jet for a video game, on par with a PS3-era model. Twenty-five minutes later the model was sitting in Blender, ready to use.
The whole job cost under a dollar. The interesting part is not the mesh quality, it is that the model drove Blender itself through a tool connection rather than handing back code to run.
K3 vs Fable 5, same globe dashboard
Kai ran the identical prompt through both models.
Kai built the same 3D globe dashboard twice, same prompt and reference, K3 on the left and Fable 5 on the right. His read: K3's UI fidelity is a touch better, and the globe has richer textures and smoother motion.
He put the cost at roughly $3 with K3 versus $15 with Fable for the same task, and said he would pick K3 for frontend work. The price gap tracks the roughly one-third pricing we found in our K3 review.
K3 vs Sol vs Fable, one game prompt
Command Code ran the same /design prompt through three flagships.
Command Code gave the same one-shot game prompt to Kimi K3, Fable 5, and GPT-5.6 Sol, then scored the results on pace, design sense, and gameplay feel. K3 came out on top of their ranking, at 9.5 out of 10 for roughly three cents a run.
Fable 5 scored 7.5 at 38 cents and Sol 7 at 11 cents in their test. Both of those, they noted, played too fast. Treat any single-run scoring as one builder's opinion, not a benchmark.
Running K3 inside Codex
Melvyn wired K3 in next to the Cursor model list.
Because K3 ships with an open API and open weights, it slots into tools that were never built for it. Melvyn set it up inside Codex, sitting next to the usual Cursor models, and called it full Codex power unlocked.
This is the quiet advantage of an open model. You do not wait for your favorite tool to add official support.
What K3 is good at, one week in
What six unrelated builds have in common.
Frontend and 3D work is where K3 keeps showing up. The Google Maps clone, the globe dashboards, and the Blender jet are all visual, spatial tasks, and in the head-to-heads builders kept giving K3 the edge on UI fidelity.
Cost comes up in nearly every post: $3 versus $15, three cents versus thirty-eight, under a dollar for a 3D asset. At a third of Fable 5's price, with open weights on top, K3 makes some tasks cheap enough that people stop rationing them.
The security demo is the most consequential item here and the least confirmed. Real bugs, real patches from Redis, and a credible researcher, but the headline number about how fast an AI did it on its own has not been checked by anyone outside the person who posted it.
For the full picture on where K3 fits, read our Kimi K3 review, compare it against the field in the best open source coding model guide, or find your own match with the 2-minute model picker.
Founder of Spectrum AI Labs — testing AI tools and models, and writing up what actually ships.
More about Paras →Stay ahead of the AI curve
We test new AI tools every week and share honest results. Join our newsletter.
