Kimi K3
Day-0 end-to-end adaptation completed on cPilot Runtime. Linux x86-64 binary preview is being packaged.
Official model card ↗Public binary research preview
Run a 2.8-trillion-parameter model locally.
K3 Flight runs Kimi K3 through a CPU-only local inference path with approximately 55GB measured runtime memory in the maintainer reference run.
NOT COMPRESSION. The checkpoint stays on storage. ~55GB describes the preliminary maintainer runtime working set, not a universal minimum.
cPilot managed flight
2.8Tparameters
LINUXx86-64
~55GBmeasured RAM
CPUonly
01 / The proof
The impossible-looking numbers are different kinds of numbers.
Total checkpoint size describes the model on storage. Runtime memory describes what the machine carries at a particular moment. They are related, but they are not the same measurement.
The checkpoint stays whole.
This is not a distilled model and the runtime does not turn 929GB into 55GB. K3 Flight keeps the complete Q2_K checkpoint available on local storage while cPilot Runtime manages the active execution window.
The result is about feasibility, not a datacenter throughput record.
The public recording will only be added when it is captured from the same release binary used for the published measurements. No simulated terminal is shown here.
Reference benchmark
Preliminary maintainer run| Model | Kimi K3 Q2_K |
|---|---|
| Total parameters | 2.8T |
| Model files | ~929GB |
| Backend | CPU-only |
| Runtime memory | ~55GB |
| Prefill | ~1 token/s |
| Decode | ~0.8 token/s |
The exact CPU and SSD models are not disclosed in this preview. Results vary with CPU capability, local storage bandwidth, context length, prompt shape and runtime version. ~55GB is a measured result, not a guaranteed minimum.
02 / The flight path
The model does not need to fit in memory. It needs a flight plan.
Kimi K3 activates 16 of 896 experts per token. cPilot Runtime turns that model sparsity into a managed execution path across local storage, memory and CPU.
Public architecture view · select a stage
One coordinated path.
Select a stage to inspect the public boundary.
Complete checkpoint on storage
The approximately 929GB Kimi K3 Q2_K checkpoint stays whole on local storage. It is not reduced to the runtime memory figure.
[LOCKED]Exact placement decisions, cache rules, prediction methods, weight layouts, memory thresholds, scheduling policies and optimized kernels remain proprietary.
03 / Model runway
K3 is the first flight, not the destination.
K3 Flight is a public validation surface for cPilot Runtime. Future cards describe our adaptation roadmap, not completed support or partnerships.
DeepSeek V4
Adaptation is planned. Performance, memory requirements and release timing will only be published after verification.
Official release ↗GLM 5.2
Adaptation is planned. Public claims will follow a reproducible cPilot Runtime run on the target release.
Official model page ↗Qwen 3.8
On the runway pending publicly downloadable weights. This does not claim current support or a confirmed delivery date.
Watch for updates ↗Model names belong to their respective owners. K3 Flight, cPilot Runtime and ONE-TOKEN are not affiliated with, sponsored by or endorsed by Moonshot AI, DeepSeek, Z.ai or Alibaba/Qwen.
04 / Community hardware map
Every honest run moves the boundary.
The first row is a maintainer result. Independent successes, slow runs and controlled failures will become the public hardware matrix.
| Source | Memory | Backend | Prefill | Decode |
|---|---|---|---|---|
| Maintainer reference | ~55GB | CPU-only | ~1 tok/s | ~0.8 tok/s |
05 / Read before flight
Frequently asked questions.
Did you compress 929GB into 55GB?
No. The model files remain approximately 929GB on storage. Around 55GB is the preliminary runtime memory observed in the maintainer reference run.
Is this a smaller or distilled Kimi model?
No. The reference run uses the complete Kimi K3 Q2_K checkpoint. Q2_K is a quantized representation and is not numerically identical to the original higher-precision checkpoint.
Why can a 2.8T model run this way?
Kimi K3 activates 16 of 896 experts per token. cPilot Runtime manages the active execution window rather than requiring the entire checkpoint to remain resident in memory.
Is ~55GB the minimum RAM requirement?
No. It is one preliminary measured result. The preview recommends at least 64GB of system memory, and results depend strongly on CPU, SSD, context and runtime version.
Is the cPilot Runtime source public?
No. This repository publishes a runnable binary research preview, documentation, checksums and community measurements. The inference engine implementation remains proprietary.
Does K3 Flight distribute the Kimi K3 weights?
No. The model is downloaded separately and remains governed by the Kimi K3 License. The K3 Flight binary has its own license.
If this changed what you thought could run locally, star the repo.
Watch the flight for the reproducible Linux preview, verified demo, new model adaptations and the growing community hardware map.