What's this about?
Chinese lab Moonshot released Kimi K3 — a 2.8-trillion-parameter model with built-in image understanding and a 1-million-token context window — and, crucially, published its weights for anyone to download. Independent evaluator Artificial Analysis scored it 57 on its Intelligence Index, just two points behind the best-configured GPT-5.6 Sol, making K3 the first open-weight model to sit essentially level with the closed frontier. A dozen cloud providers went live with it the same day.
What exactly is an "open-weight model"?
A model's weights are the billions (here, trillions) of learned numbers that are the AI — the recipe itself, not just the finished dish. "Open-weight" means Moonshot published that recipe: anyone can download K3 and run it on their own hardware, unlike GPT, Claude, or Gemini, which you can only rent through an API. Until now, open models trailed the closed leaders by a clear margin; K3 is the first to nearly erase that gap.
What's the community saying?
- "Historic — for the first time an open-weights LLM is right at the top", tempered by "we won't be able to run this ourselves" (it needs ~1.5TB of VRAM even quantized).
- The real economic tell: median third-party pricing will finally reveal what serving a 3T model costs, and competition is expected to crush it — GLM's price fell ~45% in six weeks.
- Practical scramble: can Hugging Face even serve the download rush? Will unsloth/Qwen-style distillations shrink it to consumer-runnable 20B/200B versions?
- Skepticism about the giveaway ("strange communists giving away such an expensive model") and requests for censorship / political-bias testing.
- MoE serving economics: mxfp4-native weights make it a benchmark case for what 3T-class hosting requires.