Moonshot Launches Kimi K3: Open-Source Multimodal Model Now Available on Modal
Key Takeaways
- ▸Kimi K3 is the strongest open model on public leaderboards—2.8T parameters, 1M token context, native vision—placing 4th overall against closed-source frontier models
- ▸Running at 460 tokens/sec on Modal with custom DFlash speculation, delivering 360% faster interactivity (100 → 460 tok/s) and 88% higher throughput
- ▸Moonshot's architecture innovations (Delta Attention, Attention Residuals, rebalanced expert parallelism) achieve 2.5x better scaling efficiency than K2
Summary
Moonshot has released Kimi K3, a 2.8 trillion parameter multimodal model with a 1 million token context window, marking a major milestone for open-source AI. The model is available immediately on Modal's platform, running at an impressive 460 tokens per second thanks to a custom DFlash speculator trained specifically for K3's architecture. K3 ranks as the strongest open model on public intelligence indexes and fourth overall, competing directly with closed-source frontier models.
The release represents years of engineering work to make a 3-trillion-parameter model practical and deployable at scale. Moonshot invested heavily in quantization-aware training from the SFT stage onward, novel attention mechanisms like Delta Attention and Attention Residuals, and expert parallelism optimization—achieving 2.5x better scaling efficiency than its predecessor K2. The company even contributed a new prefix caching implementation to vLLM to overcome architectural challenges. Modal's day-zero partnership enables immediate deployment with token-based pricing, featuring 360% faster interactivity and 88% higher throughput via the custom-trained speculator, with $30 in monthly free compute credits.
- Day-zero partnership with Modal and vLLM makes frontier-scale inference accessible via token-based pricing with $30/month free tier on Shared API
Editorial Opinion
Kimi K3 represents a critical inflection point for open-source AI—it's rare to see a model of this scale match closed-source competitors on quality while remaining practical to deploy across diverse hardware. Moonshot's investment in infrastructure engineering (quantization-aware training, novel attention mechanisms, prefix caching improvements) proves that infrastructure discipline is as important as model scale, setting a new standard for open model releases. Modal's day-zero custom speculation support demonstrates the platform's evolution into essential frontier AI infrastructure, while the $30/month free tier democratizes access to a genuinely competitive model. This launch signals that the open vs. closed model gap is closing rapidly, and that engineering rigor—not just parameters—defines the next generation of AI competition.
