Moonshot Promises to Give Away a Chinese AI Nearly as Smart as the Best US Models

Moonshot AI wants to hand the world, for free, a model that sits a few points behind the best paid AI America sells. The Beijing startup released Kimi K3.1 Pro late last week, and the independent testers at Artificial Analysis promptly ranked it 4th in the world. One catch. You cannot download it yet.

In its launch post, Moonshot calls K3.1 Pro its most capable 3T-class open-weight model, with a 1 million token context window and a Highspeed variant it says runs 2 to 3 times faster. It is live now in Kimi's apps and API. The weights, the part that makes "open" mean anything, are only promised. Moonshot says they will be released by Nov 4. Until a file lands, this is a hosted model with a pledge attached.

The pledge matters because of what Artificial Analysis found. Its independent evaluation scores K3.1 Pro at 55 on its Intelligence Index, an 11 point jump over Kimi K3. Only three models rank higher: Anthropic's Claude Opus 5.5 at 58, Claude Sonnet 5.5 at 56 and Meta's Muse Wonder Full at 56. If the weights ship, the firm says, K3.1 Pro becomes the top open-weights model, and the gap between the best open and best proprietary systems on its index shrinks from 12 points to 3.

Three points. Over the summer, this site argued that a handful of index points separating frontier models was less meaningful than the price gap behind them, back when Kimi K3 was the Chinese contender. K3.1 Pro pushes that argument harder.

Two sets of coding numbers

There are two sets of coding claims, from different places. Moonshot's own, self-reported figures put K3.1 Pro at 65.7% on Terminal-Bench 4.0, ahead of OpenAI's GPT-6 Astra at 58.2% and just behind Claude Opus 5.5 at 66.4%. Moonshot pulled rival scores from their blogs or the official leaderboard.

Artificial Analysis ran its own test and got a higher number. It measured 67% on Terminal-Bench 4.0, which it says is the highest score of any model it has tested, up from a dismal 13% for Kimi K3. It also clocked K3.1 Pro 3rd on GDPval-AA, its real-world knowledge work test, at 1,846 Elo, up 312 from K3 and trailing only Opus 5.5 and Muse Wonder Full. Gains elsewhere were modest, and its measured hallucination rate fell only from 53% to 44%.

Caixin reports that K3.1 Pro landed 8th on Code Arena after launch, behind frontier models from Anthropic, OpenAI and Meta. Its predecessor K3 had topped that ranking in July, the first Chinese model to do so.

Then there is the bill. Moonshot charges $2 per million input tokens and $10 per million output tokens, 33% less than K3 on every rate, according to Artificial Analysis. The savings get eaten partly by chattiness. The model used 14% more output tokens per task than K3, leaving cost per task at roughly $2. It is also no sprinter. Artificial Analysis clocks it at about 52 output tokens per second and sums it up bluntly.

"It's also notably slow and very verbose." Artificial Analysis

Moonshot spent an unusual share of its launch post on cybersecurity. It says it stripped cyber-related data from training for the open release, and that an unfiltered checkpoint solved 90.5% of its internal exploit tasks versus 81.0% for the shipped model. Company figures on a company benchmark, but the principle is plain.

"We aim not to release open weights that would extend the absolute frontier of cyber capability." Moonshot AI

The business stakes are not small. Caixin reports Moonshot closed a Series F at a $35 billion post-money valuation led by China's national AI fund, began raising a pre-IPO round at $50 billion, and got K3 onto Amazon Bedrock. A free near-frontier model is a headache for American labs charging a premium for a few extra points.

All of it hinges on Nov 4. If the weights arrive, the open-versus-closed gap is the narrowest it has looked on Artificial Analysis' index. If they slip, K3.1 Pro is just another strong, cheap, slow API. Moonshot has made the promise. Now it has to ship.