Kimi K3 with 2.8 Trillion Parameters Surpasses Opus 4.8 in Some Tests, Triggers Nasdaq Volatility

Moonshot released the Kimi K3 model in July 2026, with a parameter count of 2.8 trillion and a context window supporting 1 million tokens, available via open-weight API and developer documentation.

In some programming and general agent benchmarks, the model scored higher than Claude Opus 4.8, but overall it still lags behind Claude Fable 5 and GPT-5.6 Sol. In the officially published GDPval-AA v2 knowledge work test, Kimi K3 scored 1687 points, surpassing Claude Opus 4.8 Max's 1600 points.

Technical Approach and Cost Structure

Kimi K3 adopts the Kimi Delta Attention and Attention Residuals architecture, with an MoE design featuring 896 expert networks, of which only 16 are activated per inference. The company claims this structure improves scaling efficiency by approximately 2.5 times compared to the previous generation, increases decoding speed for million-token contexts by 6.3 times, and enhances training efficiency by about 25%.

The API pricing is $3 per million input tokens ($0.3 with cache hit) and $15 per million output tokens, roughly 70% lower than Claude Fable 5. The company's annual recurring revenue has exceeded $300 million, with APIs accounting for over 70%.

Immediate Market Reaction

Following the announcement, the Bloomberg Asia Semiconductor Index fell over 6%, Nasdaq 100 futures dropped about 2%, Z.AI's Hong Kong stock fell 28%, and MiniMax Group fell 16%. Supporters argue that the Chinese model demonstrates a low-cost pathway, while critics point out its continued reliance on US basic research and distillation techniques.

Internally at Moonshot, Kimi K3 is positioned as a product competing with Anthropic's flagship model, Claude Opus 4.8.

Impact on Developer and Enterprise Selection

The open-weight release allows global developers to freely download, deploy, and modify the model, potentially rapidly expanding its ecosystem in a manner similar to DeepSeek. Enterprises can directly invoke the model via the Kimi official website, app, desktop client, or API, with the default thinking intensity set to max mode.

For scenarios requiring long-cycle software engineering, visual understanding, and multi-tool invocation, Kimi K3 offers a deployable private option, with inference costs lower than closed-source peer models.

Changes in Competitive Landscape

Previously, market consensus held that Chinese models lagged behind US counterparts by 8 to 12 months. Kimi K3's near-comparable performance in some benchmarks challenges this perception. Anthropic plans to raise the price of Claude Opus 4.8 by approximately 50% starting September 2026, further highlighting the price gap.

Upstream and downstream, chip stocks such as NVDA face short-term pressure; Chinese peers face valuation adjustment pressure; developers gain more low-cost options, and enterprise users can allocate workloads between open and closed-source models based on task complexity and budget.

Forward-Looking Assessment

If the open-weight version of Kimi K3 sees large-scale private deployment, the rate of developer ecosystem expansion will determine whether it can sustain pressure on the closed-source approach. Signals to watch include the model's consistent scores in BrowseComp and AA-Briefcase tests, as well as changes in Moonshot's next funding round valuation.