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

Summary

The episode treats Moonshot’s Kimi K3 as a serious frontier model, with the hosts arguing it belongs in a clear top tier and may even be more practical than some competitors because it rejects fewer tasks. They spend much of the discussion on deployment economics: K3 is said to be 2.8 trillion parameters, too large for a B200, and likely requiring B300/GB300/MI355X-class systems, while delays in weight release are interpreted as a way to manage capacity, let third parties prepare performant serving, or arrange licensing/hosting partners. Pricing is another major theme, with K3 discussed at $3 per million input tokens and $15 per million output tokens—well above Moonshot’s earlier pricing—leading to debate about margins, customer segmentation, and whether there is much room to raise prices further. The hosts also broaden into geopolitics and strategy, arguing that government restrictions, Chinese domestic accelerators, and enterprise governance concerns all shape how open-source frontier models are adopted. They close by framing the AI market as still early, with new use cases continuing to expand demand rather than simply cannibalizing incumbents like Anthropic or OpenAI.

Notes

• Kimi K3 is framed as a frontier/open-source contender and explicitly discussed as a possible third-best model worldwide.
• The hosts argue composite benchmark rankings are directionally useful even if imperfect.
• Their composite top tier is described as Fable, Soul 5.6, and Kimi K3, above other open and closed models.
• Google is singled out as being embarrassed by Kimi K3’s ranking.
• Moonshot itself says K3 still trails Fable 5 and GPT 5.6 in user experience.
• In personal use, K3 is described as good but noticeably slow.
• The host says many current models are already good enough for routine work and hard to break with ordinary tasks.
• One speaker argues K3 may even be the second-best practical model because it is less annoying than Fable in day-to-day use.
• The speaker distinguishes between API use and consumer-product use, saying the API avoids rejection while web and coding products hit limits.
• Moonshot is said to lack enough GPUs to serve current demand.
• The hosts note that open weights historically solved capacity shortages by letting third parties serve demand.
• A ten-day delay before releasing weights is interpreted as a way to let external hosts prepare performant serving.
• They worry that an immediate drop could leave users with only ~20 tokens per second and damage the brand.
• Another theory is that Moonshot is negotiating with hosting/licensing partners to absorb incremental capacity.
• K3 is described as a 2.8 trillion parameter model.
• The hosts say K3 does not fit in a B200.
• Serving is said to require B300, GB300, or MI355X hardware in a single 8-way HGX server.
• Multi-node pipeline parallelism is possible, but the hosts say it would hurt performance.
• The hosts conclude that only the latest chips will be able to serve K3 well.
• Frontier performance at 2.8T parameters is taken as evidence that closed-source frontier models may be even larger than public estimates.
• They infer K3 is likely in the same size range as closed frontier models, maybe not much smaller and possibly slightly larger.
• They argue closed-source labs must have very large margins if they can serve frontier models at current prices.
• The token pricing is compared favorably to other frontier providers as evidence that API-tok economics can be excellent.
• Moonshot’s previous model was much cheaper, so K3 implies a more than 3x price increase.
• The hosts think Moonshot has limited room to push prices higher.
• Cost-sensitive users are expected to choose cheaper tiers once a model gets expensive enough.
• The market is split between heavy users who burn tokens freely and price-sensitive users with small weekly budgets.
• Those cost-sensitive users are expected to prefer low-priced models that are already good enough for everyday tasks.
• The hosts think K3’s likely audience is open-source enthusiasts more than conservative enterprises.
• K3 is treated as a new base model with a fully new architecture rather than a continuation of K2.5.
• They say the model appears complete in usage and does not show obvious raw-model failures.
• The hosts expect two or three post-training updates over the next few months.
• They expect pricing to stay stable because new hardware will not arrive quickly.
• They say current pricing for Miniax and GLM already pushes what a 1T to 1.5T model can be served for without awful margins.
• The hosts frame the biggest strategic question as whether open-source frontier models can fully close the gap to closed models.
• One speaker claims the gap closed largely because U.S. restrictions limited access to Anthropic’s best models.
• Later discussion says K3’s blog mentions quantization during the SFT stage.
• The model reportedly uses MXFP4 weights and MXFP8 activations for broad hardware compatibility.
• They connect K3 to China’s accelerator ecosystem and say frontier-model work is a national priority.
• The speakers say Moonshot should now be treated as a frontier lab.
• They argue U.S. policy can shape frontier access by restricting some labs and delaying public releases.
• They say model harnesses are part of the product, not just a wrapper.
• Usability details like remote SSH install, keystrokes, and command editing affect model choice and spend.
• The hosts suggest outcome-based pricing could produce extremely high margins.
• They argue frontier labs still have many open use cases beyond coding.
• Audio, video, deep research, robotics, and world models are named as remaining frontier opportunities.
• The speakers say day-to-day differences between frontier modes are getting harder to notice.
• Top-end users are heavy enough that they can drive substantial revenue even if capability differences narrow.
• The episode says K3 is unlikely to create net-new ARR acceleration at Anthropic or OpenAI.
• Existing users are still finding new high-ROI use cases every day.
• The hosts think those users will default to strong frontier models to unlock more value.
• They say the market is still early and adoption will keep ramping fast.
• The conference anecdote is used to argue that the AI ecosystem remains underpenetrated.
• The episode ends with a joking stock-market aside and an explicit investment disclaimer.