MODEL · MOONSHOT AI · 2.78T TOTAL / ~50B ACTIVE PER TOKEN (SPARSE MOE); NATIVE VISION VIA MOONVIT-V2
Kimi K3 (2.8T-A50B)
The largest open-weight model ever released, and the resolution of a promise we deliberately did not take on faith. Moonshot launched K3 as an API-first flagship on July 16 and committed to publishing weights by July 27; the weights landed on the 27th. It is a 2.78-trillion-parameter Mixture-of-Experts with roughly 50B active per token, a 1M-token context, native vision, and a hybrid-attention design (Kimi Delta Attention plus Attention Residuals). Artificial Analysis scored it at 57 on its Intelligence Index — Opus-4.8 / GPT-5.5 territory. The shipped repo is 96 safetensors shards totalling ~594 GB, natively MXFP4 rather than BF16.
License: Custom "kimi-k3" licence (HF-tagged `other`; Modified-MIT style — NOT classic OSI). Read it before commercial use. · Context: 1,048,576 tokens (1M) · Released: July 16, 2026 (API); open weights published July 27, 2026, exactly as promised
The decision in five lines
- The call
- Skip for local
- Best for
- Local evaluation and family reference
- Runs on
- No planner hardware fits at default quant — see model card
- Watch out
- This is a datacenter model: ~594 GB as shipped, and even aggressive community quantization leaves it hundreds of gigabytes past the largest box on our hardware list.
- Evidence
- Estimated
- 2.78T total
- PARAMETERS
- MOE
- TYPE
- 1048576
- CONTEXT
- Not locally runnable — ~594 GB of shipped weights at native MXFP4
- VRAM AT Q4
Where we recommend this
This model isn’t currently in an active planner slot. See the runner notes below if you’re running it anyway.
The call
The largest open-weight model ever released, and the resolution of a promise we deliberately did not take on faith. Moonshot launched K3 as an API-first flagship on July 16 and committed to publishing weights by July 27; the weights landed on the 27th. It is a 2.78-trillion-parameter Mixture-of-Experts with roughly 50B active per token, a 1M-token context, native vision, and a hybrid-attention design (Kimi Delta Attention plus Attention Residuals). Artificial Analysis scored it at 57 on its Intelligence Index — Opus-4.8 / GPT-5.5 territory. The shipped repo is 96 safetensors shards totalling ~594 GB, natively MXFP4 rather than BF16.
When not to use: Any local deployment. This is a datacenter model: ~594 GB as shipped, and even aggressive community quantization leaves it hundreds of gigabytes past the largest box on our hardware list. There is no configuration of consumer or prosumer hardware that runs it. Also read the licence before commercial work — it is a custom Moonshot licence, not MIT or Apache, despite frequently being described as "open source". Reasoning is always-on (`reasoning_effort` accepts only `max` at launch), so hosted output-token bills run structurally high.
Runner notes
No official GGUF or MLX from Moonshot. Community quantizations exist (`unsloth/Kimi-K3-GGUF`, `RedHatAI/Kimi-K3-FP8-BLOCK`, `RedHatAI/Kimi-K3-NVFP4`) but none bring it within reach of a single machine. Realistically you consume this hosted: `kimi-k3` on api.moonshot.ai at $3.00 per 1M input on a cache miss, $0.30 on a cache hit, $15.00 output. We list it as a frontier reference point, not a pick — the reason it matters here is that the open-weight frontier is now within one step of the closed frontier, and it keeps being Chinese labs closing the gap.
Next step
Find-by-model — see what hardware runs this→