FactStream

Release and Performance Evaluation of Moonshot AI's Kimi K3 Model

Occurred 16 July 2026

Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight model, on July 16, 2026. While the model topped global coding benchmarks and demonstrated architectural innovations, independent assessments found performance gaps in cyber capabilities and a high hallucination rate, even as the company moves toward a potential $50 billion valuation and Hong Kong IPO.

Aggregated from 12 sources · Updated 26 Aug 2026, 03:07 UTC (UTC)

Share

Coverage Balance

12 sources
Center 100%

Blindspot Alert: Center Gap

This event is primarily covered by one side of the political spectrum. Niche or counter-narrative facts may be underrepresented.

Facts (10)

  • Established

    Moonshot AI released Kimi K3 on July 16, 2026, as an open-weight model with 2.8 trillion parameters and a mixture-of-experts (MoE) architecture.

  • Established

    Kimi K3 utilizes a 1-million-token context window and architectural innovations including Kimi Delta Attention (KDA) and Attention Residuals.

  • Established

    Kimi K3 topped the Arena Frontend Code benchmark with 1,679 points, outperforming Claude Fable 5 and GPT-5.6 Sol.

  • Established

    Moonshot AI set API pricing for Kimi K3 at $3 per million input tokens and $15 per million output tokens, though cached inputs start at $0.30 per million.

  • Established

    Moonshot AI reached a valuation of $31.5 billion in July 2026 and is negotiating a pre-IPO round targeting a $50 billion valuation.

  • Established

    Moonshot AI is dismantling its offshore VIE structure to prepare for a potential listing on the Hong Kong Stock Exchange within six months.

  • Emerging

    Independent testing by Artificial Analysis reported Kimi K3 has a hallucination rate of approximately 51%.

  • EmergingNiche Signal

    A joint assessment by UK AISI and CAISI found Kimi K3's cyber capabilities performed significantly below frontier U.S. models, reaching 17 steps in a 32-step network attack compared to 28.5 steps for top U.S. models.

  • EmergingNiche Signal

    Kimi K3's safeguards did not prevent the model from assisting with agentic cyber exploit development or offensive operations during UK AISI/CAISI evaluations.

  • EmergingNiche Signal

    SpaceXAI's Grok 4.6 model, released in August 2026, surpassed Kimi K3's score on the Artificial Analysis Intelligence Index.

We'll extract atomic facts from your source and add them below for cross-checking. Single-source facts surface as Emerging or Unverified.