Kimi K3 Hits A Wall Three Days After Launch

Intelligence Brief

Moonshot AI paused new signups for Kimi K3 on Sunday, less than a week after launch. Demand pushed the system close to its limits in 48 hours. Existing users keep full access. New users wait while Moonshot adds server capacity in batches. The model is the largest open weight system released so far, at 2.8 trillion parameters, and it has already topped a major coding leaderboard.

Story Breakdown

Moonshot AI launched Kimi K3 last week. Within 48 hours the company said demand had pushed close to the limits of its systems. So on Sunday night, on X, it announced a pause on new signups. Current account holders are not affected. New users are stuck waiting until Moonshot rolls out more capacity, which it says will happen gradually.

This is not a one off story. Chinese labs keep shipping models strong enough to grab global attention, then find out they don't have the compute to handle the crowd that follows. Kimi K3 is a 2.8 trillion parameter model, the largest open weight release to date. According to the South China Morning Post, it has beaten GPT 5.6 Sol and Claude Fable 5 on some benchmarks, and it topped the Arena leaderboard for front end coding right after release.

Lian Jye Su, chief analyst at Omdia, says new releases usually bring a spike that strains infrastructure. His read is that Moonshot likely underestimated how popular K3 would get. He also points out the model is unusually heavy on compute, which makes serving it at scale both hard and expensive.

The release has already put pressure on US tech stocks. Investors are watching a familiar worry play out again, that cheap and capable Chinese models could squeeze the pricing power of American AI companies. And this is happening even with US export restrictions still limiting China's access to the most advanced chips.

Kimi K3 arrives in a crowded field. DeepSeek's V4 and its original 2025 release already reset how the world sees Chinese AI labs. Over the weekend Alibaba previewed Qwen 3.8 Max, a 2.4 trillion parameter model the company itself claims ranks second only to Fable 5, though no independent benchmarks back that up yet. Last month Zhipu, known as Z.ai, released GLM 5.2, which has already seen fast uptake outside China.

Superintelligence used to be locked inside billion-dollar quant firms whose algorithms quietly took advantage of everyone else. 

Co-Invest puts it right in your chat window. Analyze markets, manage risk, and execute trades, all inside Claude and ChatGPT. 

The institutions built the game, Co-Invest gives you a way to beat them.

Trade with Co-Invest Today

Strategic Perspective

The real story here isn't the model. It's the mismatch between how fast these labs can train something impressive and how slowly they can actually serve it. Open weight strategy wins headlines and adoption fast, but running a 2.8 trillion parameter model for millions of new users is a different problem than publishing the weights.

For anyone building on these models, the lesson is practical. Don't assume the newest, most benchmarked Chinese model is stable enough for production traffic on day one. Capacity crunches like this one are becoming part of the release cycle, not an exception to it.

Watch the next 30 days. If Moonshot reopens signups smoothly, it strengthens the case that Chinese labs can now scale infrastructure almost as fast as they scale ambition. If the freeze drags on, it tells you the compute gap is still the real bottleneck, no matter how good the model is on paper.

Why Kimi K3 Stands Out

2.8 trillion parameters, frontier scores, and a price tag that undercuts most of the competition

Kimi K3 launched on July 16 and it's already forcing people to pay attention. It's a 2.8 trillion parameter model, the largest open weight system released so far, with a full 1 million token context window and native vision built in. Moonshot has promised full public weights by July 27.

On price it charges 3 dollars per million input tokens and 15 dollars per million output tokens. Cached input drops to just 30 cents. That's the same rate as Claude Sonnet 5's standard pricing, but it comes in cheaper than Claude Opus 4.8 at 5 and 25 dollars, and cheaper than GPT 5.6 at 5 and 30 dollars. It costs roughly three to four times more than its own predecessor, K2.6, and it's well above open rivals like DeepSeek V4 and GLM 5.2.

The performance backs up the price. Artificial Analysis ranks K3 fourth out of 189 models on its Intelligence Index, just ahead of Claude Opus 4.8. It topped the Arena leaderboard for front end coding and design work. It's also strong at long codebase work, tool use, and iterating against logs, tests, and screenshots, which is where a lot of coding agents actually live day to day.

There's a catch worth knowing before you commit. K3 only runs at maximum reasoning effort right now, so there's no lighter, cheaper mode for simple questions. Every request runs a full thinking pass before it answers, and that thinking gets billed as output, which can make casual use feel more expensive than the sticker price suggests. It also tends to run more verbose and a bit slower than peers in its price class, and its web search feature is currently marked unstable.

For teams doing heavy coding or long document work, K3 lands in a strong spot. It's frontier level intelligence at Sonnet level pricing, undercutting Opus and GPT 5.6 while staying close to their scores. For simple daily chat, the always on reasoning cost is the thing to watch before you switch.

Most AI news is a waste of time. The Future Today is a daily 5-minute read focused on what actually matters.

You’ll get exclusive interviews with the CEOs, researchers, and builders shaping AI. Plus top stories broken down simply and practical advice you can immediately apply.

Read the newsletter trusted by teams at NVIDIA, Google, Anthropic, Meta, Dell, and Salesforce.

Join Free Today →