Kimi K3 goes live at 2.8 trillion parameters: Moonshot ships the biggest open-weight frontier bet yet

Moonshot takes Kimi K3 to GA with a striking figure: 2.8 trillion parameters, the largest open-weight architecture announced to date.

In plain terms. Moonshot AI has moved Kimi K3 from the preview phase to live, with a displayed figure of 2.8 trillion parameters. On paper, it's the largest open-weight architecture to date. Remains to be seen what it's worth outside of in-house benchmarks.

Context

The thread opens at the beginning of the month: Moonshot floods Weibo and HN with benchmark videos even before the release (thread kimi-k3-launch, post #1164), then goes GA on July 16, 2026 on kimi.com/en with code docs. Techinasia relays on July 17 the official figure: 2.8T parameters, without further details on activation per token or sparsity ratio, which is the question that matters for a MoE of this size.

The data

• Announced size: 2.8 trillion parameters (source: Techinasia, July 17, 2026). That's ~2× GPT-4 according to commonly cited leaks, and clearly above the public DeepSeek variants (671B for V3).
• Availability: GA via kimi.com/en and documented API.
• Ecosystem: coordinated announcements on HN, Weibo, dev channels - confirming the "flood then ship" strategy described in #1164.

Analysis

Two things to separate. The figure: 2.8T is as much a marketing signal as an architectural one. In MoE, what matters is the number of parameters activated per token, the expert topology, and the effective inference cost. As long as Moonshot doesn't publish these metrics, "2.8T" remains an ad. Availability: going GA a few days after the preview is a deliberately aggressive pace - it aims to capture the attention window before third-party benchmarks (LMArena, SWE-bench, MMLU-Pro) render their verdict.

Under the hood

What to look for in the coming days:

1. Active parameters per token - without this figure, "2.8T total" says nothing about the inference cost.
2. API pricing vs Claude/GPT on long tasks (context, tool-calling).
3. Reproducibility of in-house benchmarks on SWE-bench Verified and LiveCodeBench.
4. Latency in tool-use - raw size penalizes if orchestration is not careful.

So what

If the figure holds (reasonable activation, competitive pricing, consistent third-party benchmarks), Moonshot signs the first open-weight model that directly compares to the GPT-5/Claude 4.7 generation - and does so from China, despite the compute squeeze. If the figure is marketing wrapping around a lightly activated MoE, K3 joins the long list of "world's largest models" forgotten in three weeks. We'll know by the end of July.