Kimi K3 Is the Next DeepSeek Event

Moonshot AI released Kimi K3 last week, and it's the first open weight model since the original DeepSeek moment that genuinely plays in the frontier league. Not "impressive for an open model". Actually competitive compared to even the big boys.

One disclaimer before the opinions start: The performance numbers here are not mine, and instead come from Moonshot's own release material, Artificial Analysis' independent testing, and hands-on runs by people like Theo at t3.gg.

What it actually is

In short, it is a really impressive model. 2.8 trillion parameters, which makes it the largest open-weight model ever released and almost three times the size of its predecessor. Native visual capabilities, a 1 million token context window, and clearly built for long-horizon agentic work. Moonshot's own benchmarks put it ahead of Claude Opus 4.8 and GPT-5.5 in several areas, coding, visual understanding, and long-running tasks among them, while still trailing the latest and greatest, such as Claude Fable 5 and GPT-5.6 Sol overall. Early independent testing, including Artificial Analysis' intelligence index among others, seems to mostly confirm that assessment.

That's something I haven't been able to write about an open-weight model in a while. DeepSeek had its stock-crushing moment, but ever since, open-weight providers haven't been able to offer capabilities actually close to frontier models. There have been impressive ones, even Kimi K2 was one, but they were always impressive as open weight models and considerably behind the big ones. And in the end, they became more like blips in time when one specific model was briefly a decent option, then the frontier moved on, and the blip faded. In that context, Kimi K3 starts from a higher point than any of them before.

Surprisingly good at visual work

One genuinely unexpected bit: K3 is strong at graphical work. UI design, front-end code, even 3D scenes. Within a day of release, it took the number one spot on Arena's Frontend Code leaderboard, ahead of Fable 5 and GPT-5.6 Sol in blind human evaluations. People have been running it through visual coding tasks too, with results landing somewhere around "slightly behind Claude, ahead of OpenAI". The native vision side does real work here: the model iterates against screenshots of its own output, sees what it built, and fixes it. It can even handle 3D work better than the competition.

No guardrails

Then the part that makes this release potentially interesting: K3 ships with essentially none of the usage restrictions the US labs build in.

Anthropic, for example, instructs its models not to help competitors with AI research. Somewhat understandable from a business perspective. Also a strange world where the tool gets to decide whose research is allowed. Fable also regularly refuses security research, even on your own project. Point it at a codebase and ask it to find vulnerabilities, and it declines, because the same request would work against any open source code as well. Again, understandable. But also regularly infuriating when it's your own production system you're trying to harden.

K3 doesn’t have any of this. Security research, GPU kernel optimization, even chip design, the areas where the big labs have drawn their lines. For security work on your own systems, it's arguably the most capable model that will actually agree to do it.

The catches: speed, data, hardware

The biggest one is speed. K3 currently only runs at maximum reasoning effort, and Artificial Analysis measured it at around 60 output tokens per second, below its class average, while still producing lots of output tokens. When comparing against GPT-5.6: half the price per token, double the tokens, so the cost mostly evens out while the work finishes lot slower. Artificial Analysis' cost-per-task numbers land in the same place: $0.94 per Intelligence Index task, right next to GPT-5.6 Sol's $1.04, roughly half of Opus 4.8's $1.80 and even smaller fraction of Fable 5's $2.75. So per finished task, it undercuts the Anthropic models and matches OpenAI. You just wait longer for the result. Cheaper on the wallet, not on the clock.

Second catch: right now the only way to use it is through Moonshot's own apps and API, which run in China. Your prompts, your source code, everything goes there. For hobby projects, fine. For client work, probably no.

The weights are promised for July 27, per Moonshot's own release notes. Even then, you won't be running this at home. At 2.8 trillion parameters, the model lands north of a terabyte of memory even in its native 4-bit quantized form, and Moonshot itself recommends deploying on setups with 64 or more accelerators. So realistically it's the inference providers with distributed hardware that will run it, not your workstation. The interesting question is whether the current generation of custom LLM accelerator hardware can even fit a model roughly three times bigger than what those operators have seen before. If it can, the speed problem might solve itself. If it can't, K3 stays slow until the hardware catches up.

What this actually means

If you already pay for a Claude or OpenAI subscription, there's not necessarily much reason to reach for K3 day to day. Outside the restricted areas, the big-boy models you already have available are faster and mostly better. So, even with the hype being mostly true, nothing about your current setup became obsolete last week. K3 is worth a test run out of curiosity and worth bookmarking for the security work, but there's no need to switch.

Applications are a different story. When your software uses models, that happens through an API and not subscriptions. Instead you pay the full per-token prices, and at ~$1 per finished task with no usage restrictions, K3 is definitely competing head-on with Fable and Sol in that space. And not as a budget alternative, but as a genuine full replacement for a good chunk of workloads.

Until now, open-weight models have mostly lived in the cheap end of application work: you use the frontier models to build the software, and the software itself then might run smaller open models to keep operating costs down, while the more complicated frontier-level tasks inside the applications stayed with the big labs. K3 is the first open-weight model that can credibly take those too.

And that just might be the thing that gives it legs. Application traffic is big chunk of revenue and data, and all of that is what would give Moonshot the means to keep K3 up to date when the big labs ship their next models. Every previous open weight champion fell off exactly at that point: one good release, then silence while the frontier moved on. If Moonshot can answer with updates that keep pace, this becomes a real shift in who gets to run frontier-level intelligence. If not, it's stays as just another blip. Time will tell.

Either way, it's the most promising blip yet.