[Emergency Episode] Moonshot’s Kimi K3 has Arrived! China has a Frontier Model

Summary

This emergency episode argues that Moonshot’s Kimi K3 is a genuine frontier model and perhaps the third best in the world, based on a composite view of major benchmarks. The hosts emphasize that the model is extremely large at 2.8 trillion parameters, too big for a B200, and likely requires B300/GB300 or MI355X-class hardware for efficient single-system serving. They also discuss why Moonshot may be delaying weight release, including demand management, partner serving readiness, and possible licensing arrangements with inference providers. Pricing at roughly $3 per million input tokens and $15 per million output tokens is framed as both a major increase from Moonshot’s prior model and evidence that frontier inference can carry very high margins. The conversation broadens into architecture speculation, post-training and quantization details, Chinese domestic accelerator support, and the idea that harness quality and product layer decisions increasingly determine who captures value.

Notes

• Kimi K3 is framed as a frontier-level model and likely the third best model overall.
• Composite benchmarks are treated as directionally useful despite known limitations.
• The benchmark composite is described as including major closed and open labs, making Moonshot’s result unusually strong.
• The blog reportedly acknowledges a user-experience gap versus the best closed models.
• Jordan says K3 feels good but is very slow in personal use.
• Moonshot is seen as underprovisioned on GPUs relative to demand.
• The weights delay may be intended to give serving partners time to optimize throughput.
• Another speculation is that Moonshot is negotiating licensing or hosted-serving deals with providers like Together, Fireworks, and Nebius.
• Kimi K3 is said to have 2.8 trillion parameters.
• Jordan says the model does not fit in a B200, making inference a hardware-constrained problem.
• Single-system serving is said to require B300/GB300 or MI355X-class hardware.
• Pipeline parallelism across multiple nodes is described as a material performance penalty.
• The hosts infer that leading closed-source models may be even larger than outsiders expect.
• K3 pricing is discussed at about $3 per million input tokens and $15 per million output tokens.
• That price is compared to Sonnet, suggesting K3 is now in frontier-pricing territory.
• Moonshot’s earlier model was materially cheaper, implying a more than 3x price increase.
• The new price may already be close to the ceiling for many users.
• The hosts split users into frontier power users and cost-sensitive users.
• K3 is characterized as a new base model rather than a small incremental update.
• Jordan cites Kimi Delta attention, attention residuals, and stable latent MoE as architectural elements.
• Cursor is said not to be planning a Composer-style model built on K3.
• More post-training updates are expected over the next few months.
• Pricing is expected to hold steady for the next few months because throughput gains are not imminent.
• Jordan argues the open-vs-closed gap may have narrowed because U.S. restrictions limited access to Anthropic’s best models.
• Frontier access is portrayed as increasingly dependent on government approval.
• There is strong demand for a Western open-source model that is actually competitive.
• Moonshot’s blog is said to mention quantization during SFT and native MXFP4/MXFP8 for broad hardware compatibility.
• The hardware-compatibility angle suggests an effort to broaden deployability across systems.
• China is described as running frontier models on domestic accelerators.
• Domestic accelerator support is treated as a national priority in China.
• Jordan argues the harness is part of the product, not just the model.
• Small workflow details like SSH installability and command editing can change routing decisions.
• Users may pay frontier-model prices even for tasks a cheaper model could likely handle.
• Slack bots and computer-use tools can hide the underlying model, making harness quality and routing more important.
• Outcome-based pricing is floated as a way to support extremely high margins.
• The hosts think labs may continue pushing coding models toward RSI while withholding the strongest versions from users.
• Robotics and world models are named as paths beyond coding and knowledge work.
• Jordan says software engineers spend far less than he does on these models.
• Newer users are still discovering high-ROI use cases and may default to frontier models.
• The AI market is still early if most conference attendees have not heard of SemiAnalysis.