Moonshot AI's Kimi K2.5 Now Available on Nscale
The open-weight model turns screenshots into code, coordinates 100 agents, and beats Claude Opus on research tasks.
Moonshot AI's frontier model, Kimi K2.5, is now accessible via fully managed inference endpoints on the Nscale AI cloud platform. This is a significant deployment for a 1-trillion-parameter, open-weight model that uses a Mixture-of-Experts (MoE) architecture for efficient inference. Unlike models with vision tacked on later, Kimi K2.5 was trained natively on 15 trillion mixed visual and text tokens, processing images and text as a unified stream. This enables advanced capabilities like turning screenshots into fully functional front-end code and coordinating up to 100 parallel AI agents. On Nscale, developers gain per-request control over a 'thinking mode' for multi-step reasoning, balancing latency and quality.
Kimi K2.5 excels in three key areas, outperforming leading closed models on specific benchmarks. First, in design-to-code, it scores 76.8% on SWE-Bench Verified for engineering tasks. Second, for autonomous research, it achieves 60.6% on BrowseComp, beating Claude Opus 4.5 (37.0%). Third, for document automation, it leverages its 256K context to score 87.6% on the complex GPQA-Diamond benchmark. Running on Nscale provides predictable, token-based pricing ($0.45/$2.20 per 1M tokens), fully managed scaling, and a Prompt Workbench for testing, with strict data isolation and no training on customer data.
- 1-trillion-parameter open-weight model with Mixture-of-Experts architecture for efficient, frontier-level performance.
- Native multimodal training enables advanced design-to-code (76.8% SWE-Bench) and beats Claude Opus on research (60.6% BrowseComp).
- Available on Nscale with managed endpoints, 'thinking mode' control, and predictable pricing at $0.45/$2.20 per 1M tokens.
Why It Matters
Delivers closed-model agentic and visual reasoning performance in an open-weight format, enabling customizable, production-ready AI workflows.