Two weeks ago I wrote that GLM-5.2 had pushed open weights across the frontier line. I thought that headline would hold for at least a quarter. It held for fourteen days.
Late on Sunday evening US time — a day ahead of its own announced schedule — Moonshot AI published the full weights of Kimi K3: 2.8 trillion parameters, a million-token context window, native multimodality, and benchmark placements that would have been unthinkable for an open model a year ago. It debuts at #3 on Artificial Analysis’ Intelligence Index, behind only Claude Fable and GPT-5.6 Sol Max, #2 on the Vals AI index, and #1 outright on Frontend Code Arena. This is no longer “open weights are catching up.” This is the strongest open model ever released, sitting one seat away from the closed frontier — and anyone with the bandwidth can download it tonight.
The specs, briefly
- 2.8 trillion total parameters in a mixture-of-experts design — 16 of 896 experts active, 104 billion parameters live per token, per the released model card.
- Kimi Delta Attention (KDA) and Attention Residuals (AttnRes) — Moonshot claims a ~2.5× scaling-efficiency improvement over Kimi K2.
- 1,048,576-token context, native vision, agentic post-training.
- 1.4 TB on disk in MXFP4 four-bit precision. At 16-bit it is 5.6 TB.
That last number is the one to sit with, because it reframes what “open” actually delivers here.
The 1.4-terabyte question
Loading K3 takes roughly eighteen 80 GB accelerators before you have reserved a single byte for context. A current-generation eight-card node with 192 GB per card — about 1.5 TB of pooled memory — barely fits the weights, with almost nothing to spare for a million tokens of KV cache. The model runs natively on NVIDIA Blackwell and AMD MI400 silicon, which is a polite way of saying: it runs in data centers.
The inference economics are stranger than the storage. Because only ~50B parameters activate per token, K3’s compute cost per token resembles a mid-size model. But all 2.8 trillion parameters must stay resident in memory, so the binding constraint is memory capacity and bandwidth, not FLOPs. That is exactly the resource that is scarce and getting scarcer — I wrote about the DRAM squeeze in The AI Memory Wall, and K3 is the memory wall wearing a trench coat.
So the practical beneficiaries of this “free download” are inference providers, hyperscalers, and the handful of enterprises already operating Blackwell or MI400 racks. Together AI and Modal had hosted endpoints live on day zero. For everyone else, K3’s openness is real but mediated: you will consume it through someone’s API, negotiate someone’s hosting margins, and the sovereignty benefits — your weights, your infrastructure, your data path — accrue only if you can afford the racks. Open weights at this scale decentralize control in principle and recentralize operations in practice.
Escalation is the strategy
K3 did not arrive in a vacuum. Xi Jinping publicly committed China to open-source AI at the World AI Conference. Alibaba has already teased Qwen 3.8 at 2.4 trillion parameters, weights to follow. The open-to-closed capability gap, which sat at six to nine months for most of 2024–2025, has compressed to three to five months. Releasing frontier-adjacent weights for free is not charity; it is a deliberate strategy to commoditize the layer where US labs make their margins, and it is working well enough that Washington has started responding in kind.
Which brings us to the uncomfortable part. The White House OSTP’s Michael Kratsios accused Moonshot last week of training K3 on export-restricted NVIDIA hardware and of running large-scale distillation against US frontier models — Anthropic’s Fable among them. Moonshot disputes this, and none of it is proven publicly. But it hardly matters for the strategic picture: if distillation at scale works even partially, every closed frontier release now has a half-life, and the frontier labs know it. Expect API-side countermeasures — stricter rate governance, output watermarking, tiered access — to accelerate.
What I would actually do with this
For teams building on LLMs, K3 changes three calculations:
- Your fallback tier just got much better. If your architecture routes premium traffic to a closed frontier model and bulk traffic to an open model, the quality ceiling of that second tier jumped meaningfully. Re-run your routing evals; the split that made sense in May is stale.
- Negotiating leverage is real even if you never deploy it. A credible open alternative one benchmark seat below the frontier is the best procurement lever you will ever hold in a vendor conversation. You do not have to run K3 for K3 to save you money.
- Do not confuse downloadable with operable. If someone on your team proposes self-hosting K3 for “data sovereignty,” the correct first question is whether you have ~1.5 TB of pooled accelerator memory and someone to babysit it. For most organizations the honest sovereignty play is a European or on-shore inference provider hosting open weights under a contract you can audit — not a rack you cannot afford to idle.
Fourteen days ago the question was whether open weights could reach the frontier. Now the question is who can afford to lift them. That, increasingly, is where the moat went: not the weights, but the memory to serve them.