Moonshot AI Releases Kimi K3: 2.8T Parameter Open Model at $3/$15 Per Million Tokens

Moonshot AI has released Kimi K3, a 2.8 trillion parameter model that the company says will become the largest open-weights model when weights are released on July 27, 2026. The model is live on Kimi.com, Kimi Work, Kimi Code, and via API.

Model Specifications

• Parameters: 2.8 trillion total, with 896 experts and 16 activated (under 2% activation ratio)
• Context window: 1 million tokens
• Modalities: Native multimodal input (text + images), text output
• Pricing: $3 per 1M input tokens, $15 per 1M output tokens; cached input discounted 90% to $0.30 per 1M
• Architecture: Uses Kimi Delta Attention (KDA), Attention Residuals (AttnRes), and LatentMoE

Performance Benchmarks

According to independent testing by Artificial Analysis:

• AA Intelligence Index: 57 (comparable to Opus 4.8 and GPT-5.5, behind Fable 5 and GPT-5.6 Sol)
• GDPval v2 Elo: 1668
• AutomationBench-AA: 53% (#1 ranking)
• AA-Briefcase Elo: 1547
• Cost per task: $0.94 average (vs $1.80 for Opus 4.8)

In community Arena evaluations, K3 achieved #1 in Frontend Code Arena with 1679 points and a 76% pairwise win rate, surpassing Claude Fable 5 (63%) and GPT-5.6 Sol (58%). The model ranks #9 in Text Arena with 1486 points, jumping from #38 for its predecessor K2.6.

Technical Architecture

The model introduces three key innovations:

1. Kimi Delta Attention (KDA): Claims to enable up to 6.3x faster decoding in million-token contexts. Development reportedly started in January 2025 and took approximately 1.5 years to reach production scale.

2. Attention Residuals (AttnRes): According to Moonshot, delivers approximately 25% higher training efficiency at less than 2% additional cost.

3. LatentMoE with Stable LatentMoE: Uses 16 activated experts out of 896 total, with quantile load balancing.

Additional architecture details include per-head Muon optimization and a new activation function called SiTU (Sigmoid Tanh Unit).

Inference Performance

Early testing shows inference speeds of approximately 26-28 tokens per second via Moonshot API on OpenRouter. Moonshot recommends deployment on configurations with 64+ accelerators for optimal efficiency.

vLLM announced day-zero support for K3, with Moonshot contributing a KDA prefix caching implementation directly to the vLLM project. The contribution was necessary because KDA breaks assumptions in conventional prefix caching.

Performance Caveats

Moonshot acknowledged that despite competitive benchmark scores, K3 has a "noticeable gap in user experience" compared to Claude Fable 5 and GPT-5.6 Sol.

Artificial Analysis reported the model achieved 46% accuracy on AA-Omniscience (up from 33% on K2.6), but hallucination rates worsened to 51% from 39%.

ProgramBench author Ofir Press noted that Moonshot used averaging implementation percentage rather than counting fully working programs, which may overstate performance on coding tasks.

Cost Comparison

At 80% input / 20% output ratio, K3 costs approximately $5.40 per million tokens, compared to $9 for Opus 4.8 and $10 for GPT-5.5. Artificial Analysis measured $0.94 average cost per Intelligence Index task versus $1.04 for GPT-5.6 Sol and $1.80 for Opus 4.8.

The model consumed 132 million output tokens across the Intelligence Index tests, a 21% reduction from K2.6's 166 million tokens, while gaining 13 index points.

What This Means

Kimi K3 represents a significant scale increase for open models—if weights are released as promised on July 27, it will be more than double the size of previous open releases. The combination of competitive performance with Opus 4.8-class capabilities at significantly lower pricing ($5.40 vs $9 per million tokens blended) creates a new efficiency tier for developers.

The Frontend Code Arena #1 ranking is particularly notable as a human-preference evaluation rather than static benchmark. However, the acknowledged user experience gap versus top closed models and increased hallucination rates suggest K3 may require careful prompt engineering for production use. The 1.5-year development cycle for KDA indicates substantial research investment in long-context optimization.