Moonshot AI Releases Kimi K3, a 2.8-Trillion-Parameter Open-Weight Model That Rattles Chinese Tech and Chip Stocks

Overview

Beijing-based Moonshot AI unveiled Kimi K3 on July 16, an open-weight language model with 2.8 trillion parameters that Moonshot AI delivers as the largest open-weight AI model ever. The release, as previously reported, follows Moonshot’s $2 billion funding round in May, and it landed hard: Chinese AI rivals and global chip stocks sold off the following day as investors weighed whether frontier-level performance now requires far less compute spending than assumed. Within days, demand for the new model outpaced Moonshot’s own capacity to serve it.

What We Know

Moonshot said K3’s overall performance still trails Anthropic’s Claude Fable 5 and OpenAI’s GPT 5.6 Sol, but on benchmarks spanning coding and general-agent tasks, the company said K3 performed “competitively” with Fable 5 and “substantially outperformed” Claude Opus 4.8, GPT 5.6 Sol, and GPT 5.5. Independent analyses from Arena.ai and Vals AI suggested Kimi is competitive with flagship frontier models, according to TechCrunch, lending outside support to Moonshot’s own claims rather than leaving them as an unverified vendor assertion. For comparison, DeepSeek V4 has 1.6 trillion parameters, roughly half of K3’s total, underscoring how far Moonshot pushed model size relative to its closest domestic rival.

On pricing, K3 costs $15 per million output tokens, compared to $4.40 per million output tokens for z.ai’s GLM-5.2 and $0.87 for DeepSeek V4 — still far cheaper than Fable, which costs $50 for the same amount of output, Fortune reported.

Moonshot president Yutong Zhang said, “We knew we didn’t have the luxury to simply scale up compute”, framing the model’s efficiency-first design as a response to constraints on the chips available to Chinese AI labs. Moonshot raised $2 billion in funding in May, valuing the company at over $20 billion, and its annual recurring revenue has exceeded $200 million.

The announcement coincided with a speech from Chinese president Xi Jinping at the World AI Conference in Shanghai, TechCrunch reported. The timing amplified the market reaction: on Friday, the Nasdaq dropped about 1% as investors sold off stocks in chip companies like Nvidia. Fortune’s own tally of Friday’s moves was steeper in places — Taiwan Semiconductor Manufacturing Company fell 7%, SoftBank fell 9.0%, Nvidia fell 1.2%, Meta fell over 2.4%, and the Nasdaq 100 was down 1.0% as of 2:00pm Eastern time. Domestic rivals took the biggest hits: shares of Zhipu and MiniMax plunged 28.4% and 15.6% respectively in Hong Kong trading, while Z.ai similarly dropped 28.4% — a move Fortune separately clocked as Z.ai plunging by almost 30%. TechCrunch described the reaction as carrying echoes of DeepSeek’s open-source R1 release in January 2025, noting that “now, everything seems heightened”.

Demand for K3 quickly outstripped Moonshot’s own capacity. Over the weekend following the launch, the company paused new subscriptions, saying “Over the past 48 hours, demand has pushed close to the limits of our current capacity. Our GPUs are feeling it. We’re adding capacity as fast as we can and will reopen new subscription spots in batches”. Existing subscribers, Moonshot said, would not be affected by the pause.

What We Don’t Know

Moonshot has not published a full model card, and the company’s specific architectural claims beyond parameter count have not been independently verified through sources this article could confirm. Moonshot’s own benchmark framing — that K3 trails Fable 5 and GPT 5.6 Sol overall while beating Opus 4.8, Sol, and GPT 5.5 on select tasks — is a company-supplied comparison; the article does not have access to raw scores from an independent third-party leaderboard beyond the general corroboration TechCrunch cited. The exact day the subscription pause took effect is unclear — outlets reported it happening “on Sunday” without agreeing on the calendar date. It also remains to be seen whether the market reaction to K3 proves as lasting as the reaction to DeepSeek’s R1 release in January 2025, or whether Chinese chip and compute constraints will limit how widely Moonshot can scale access to the model once the current subscription pause lifts.

Analysis

The scale of Friday’s selloff — hitting not just Moonshot’s direct Chinese competitors but Taiwanese and Japanese suppliers and U.S. chipmakers alike — suggests investors read K3 as evidence that frontier-level performance is achievable without the level of compute spending currently priced into AI infrastructure stocks. TechCrunch’s framing of the reaction as a “second DeepSeek shock” captures that dynamic: a comparatively low-cost, open-weight release from a Chinese lab once again forced markets to reprice assumptions about the AI buildout, even as Moonshot’s own capacity crunch within days of launch showed that serving that same low-cost model at scale is its own unsolved constraint. The gap between K3’s list price and Fable’s — $15 versus $50 per million output tokens, by Fortune’s accounting — is the concrete number underpinning that repricing: it lets Moonshot undercut the leading U.S. model by more than three-to-one even as its own benchmark results describe a system still a step behind the frontier rather than past it.