Kimi K3: The open-weights escalation

What it claims

Moonshot AI released Kimi K3 on 16 July, a 2.8T-parameter MoE model, with weights promised for 27 July. The post was written before that date, so its open-weights argument is conditional on Moonshot keeping the promise. Nathan Lambert's central estimate is that the open-to-closed (and US-to-China) capability gap has shrunk from the commonly debated 6–9 months to roughly 3–5 months. The author calls K3 the closest open models have been to the frontier since DeepSeek R1. The difference is that K3 came from scaling known levers (data, algorithms, architecture, tools, environments), not from a fast pivot to a new paradigm. The author also argues that K3's quality rules out adversarial distillation from US models as more than a small contributor.

How it works

The architecture details are quoted from Moonshot's launch blog:

• Kimi Delta Attention (KDA) improves information flow across sequence length.
• It was introduced in Moonshot's earlier Kimi Linear paper.
• It is related to Gated DeltaNet, which was introduced in late 2024, builds on Mamba, and was used in Ai2's Olmo Hybrid.
• Qwen's latest models moved to a related architecture.
• Nemotron is hybrid too, but closer to Mamba.
• Attention Residuals (AttnRes) improve information flow across model depth.
• Sparser MoE activates 16 of 896 experts, paired with a "Stable LatentMoE" framework.
• Refined training and data recipes complete the changes.

Moonshot attributes an approximately 2.5× gain in overall scaling efficiency over Kimi K2 to these changes combined. The page gives no per-component ablation.

Evidence

All benchmark figures are rankings only. The page gives no scores and no evaluation settings.

The author's lab ranking by peak model puts Moonshot third:

1. Anthropic
2. OpenAI
3. Moonshot (K3)
4. SpaceXAI (Grok 4.5)
5. Zhipu (GLM 5.2, open weights)
6. Meta (Muse Spark 1.1)
7. DeepMind (Gemini Flash 3.5)
8. Alibaba (Qwen 3.7 Max)

How to use it

• The API is live.
• Moonshot paused new subscriptions to its K3 consumer access because of demand.
• Weights were scheduled for 27 July.
• The page states no licence, hosting requirements or pricing figures.

Adjacent releases to watch:

• Alibaba announced an open-weights Qwen 3.8 at 2.4T parameters. This is a break from Alibaba's habit of keeping its largest models API-only.
• DeepSeek V4 is expected to leave preview.

Limits

• Weights not yet out when written. Every open-weights conclusion assumes the 27 July release happened. The post was written about two weeks before that date. Check the current weights status and licence before relying on it.
• No scores. The benchmark claims are leaderboard positions without scores, and the page reproduces none of Moonshot's own evals.
• Efficiency is self-reported. The 2.5× figure is Moonshot's own claim against its own prior model.
• Thin cost and compute evidence. Claims about Chinese labs' capital efficiency, compute access (including skirting export controls) and emerging data vendors are hearsay. The author says there is "very limited measurement into them."
• US hosting risk. The policy section cites Axios reporting on measures the US administration has considered against Chinese labs:
• Commerce Entity List additions.
• An NSA/ONCD advisory.
• An executive order making US hosts of Chinese models guarantee security and accept liability.
• Supply-chain rules targeting Chinese open-source models.

None had been enacted at the time of writing. They are still a deployment risk if you plan to self-host K3 in the US.