Kimi K3 vs Kimi K2: What Changed?
K3 is not a small revision of K2. It expands the model from 1T to 2.8T parameters, raises the context ceiling from 128K to 1M, adds native vision, and changes both sequence attention and depth-wise information flow.
Specification snapshot
The table compares the original Kimi K2 open model with the K3 launch specification. Product variants between these generations may have different limits, so use this as an architecture-level comparison.
| Specification | Kimi K2 | Kimi K3 |
|---|---|---|
| Total parameters | 1T | 2.8T |
| Selected / total experts | 8 / 384 | 16 / 896 |
| Context window | 128K | Up to 1M |
| Native vision | No | Yes |
Architecture changes
K2 used MLA attention and a 384-expert MoE. K3 introduces Kimi Delta Attention, Attention Residuals and Stable LatentMoE. Moonshot says the combined architecture, training and data changes improve overall scaling efficiency by about 2.5 times over K2.
Attention Residuals is especially important conceptually: instead of treating layer history as a uniform accumulation, the model can retrieve representations from different depths as needed.
Should K2 users migrate?
For hosted API work, the current Kimi model list directs users toward K3 and lists the older K2 API series as retired. For self-hosted research, K2 remains useful because its weights, report and deployment guides are already available, while K3's full-weight rollout and technical report follow the launch schedule.
- Move new hosted projects to K3 unless a product constraint says otherwise.
- Start a new session rather than switching an active K2 conversation to K3.
- Benchmark your own tasks; different harnesses can change public scores materially.
Frequently asked questions
Is K3 simply a larger K2?
No. It is larger, but it also changes attention, residual information flow, MoE sparsity, quantization-aware training and multimodal capability.
Does K3 replace every K2 deployment?
Not automatically. Hosted API users have a clear migration path, while self-hosters should wait for the exact K3 weights, license and inference guidance required by their stack.
Primary sources
This guide prioritizes official Kimi and Moonshot AI materials. Product access and pricing can change, so confirm live details at the linked sources.