Databricks Adds Kimi K3: The Open-Weight Model That Rivals Frontier Proprietary AI at Half the Cost

• Total parameters: 2.8 trillion
• Active parameters: 104 billion activated per token
• Architecture: Sparse MoE with 896 experts, 16 active per token, plus 2 shared experts
• Context window: 1,048,576 tokens (1 million)
• Modalities: Text and native vision (images and video)
• Architectural innovations: Kimi Delta Attention (KDA), Attention Residuals (AttnRes), and Stable LatentMoE framework

Moonshot calls K3 "the world's first open 3T-class model" . The open weights were released on July 27, 2026, under a modified MIT license .

Artificial Analysis Intelligence Index ranking

On the independent Artificial Analysis Intelligence Index (v4.1), Kimi K3 scores 57 in its standard configuration and 60 on the "max" configuration . Its ranking is :

• #3 or #4 overall among all models tested, depending on the configuration
• First among open-weight models (next-closest open-weight: GLM-5.2 at 51)
• Behind: Claude Fable 5 (~60) and GPT-5.6 Sol (~59)
• Ahead of: Claude Opus 4.8 (~56), GPT-5.5, DeepSeek V4 Pro (44), and GLM-5.2 (51)

The index combines nine evaluations including GDPval-AA v2, Terminal-Bench 2.1, SciCode, Humanity's Last Exam, GPQA Diamond, and long-context reasoning . On specific benchmarks, Kimi K3 reached #1 on the Frontend Code Arena leaderboard and #2 on AA-Briefcase (agentic knowledge work), behind only Fable 5 .

Cost comparison to proprietary models

Kimi K3 offers a significant cost advantage over proprietary rivals :

• Cost per Intelligence Index Task: Approximately $0.94
• Roughly half the cost of Claude Opus 4.8 (max reasoning, $1.80)
• Similar cost to GPT-5.6 Sol (max, $1.04)
• Databricks reports 50–72% lower cost per task compared to proprietary leaders
• API pricing: ~$3.00 per million input tokens and ~$15.00 per million output tokens

As Databricks CEO Ali Ghodsi told CNBC, K3's combination of benchmark performance and cost efficiency makes previously cost-prohibitive agentic workloads — such as continuous coding agents or large-scale document processing — economically viable for enterprises .

Unity AI Gateway enterprise governance and cost controls

Unity AI Gateway provides a unified control plane for platform teams managing AI across the enterprise. Key capabilities applicable to Kimi K3 include :

• Fine-grained access permissions and safety guardrails enforced on model inputs and outputs
• Complete audit trails and runtime observability for all model calls
• AI spend controls — proactive budget alerts across users, workspaces, use cases, and entire accounts, combined with Unity Catalog system tables and Databricks budgets
• Rate limiting and traffic routing across models, agents, and MCP servers
• Hard spend caps that stop requests when a budget is reached
• Zero Data Retention (ZDR) for Kimi K3 on Databricks
• Unified governance across both open-weight and proprietary models through the same policy framework

During the same week, Unity AI Gateway reached General Availability, unifying spend management, rate limiting, and access control across enterprise LLMs, MCP servers, and coding agents .

Broader trend: the open-weight gap is closing rapidly

Databricks frames the Kimi K3 integration as evidence that the gap between open-weight and proprietary AI models is closing rapidly . The company notes that a year ago, the best open-weight models trailed their proprietary counterparts by a wide margin — but today, open-weight models like Kimi K3 rival proprietary leaders on benchmark performance while offering dramatically lower costs and greater flexibility .

According to Databricks, this shift enables enterprises to adopt hybrid AI strategies, mixing open-weight and proprietary models on a single governed platform. The company emphasizes that enterprises now have genuine choice — with the ability to route workloads to the most cost-effective model for each task, without sacrificing performance or governance .