Kimi K3 2.8T — VRAM, speed & local setup
Written by Jakub Rusinowski · Last updated
The full 2.8-trillion-parameter Kimi K3 MoE, activating 104B parameters per token by selecting 16 of 896 experts. Moonshot's largest and most capable model, positioned above Kimi K2.6 on agentic coding and long-horizon reasoning at release. Native text + vision with a 1M-token context. Weights went public on 26–27 July 2026 under the bespoke Kimi K3 License — open-weight, but not an OSI-approved open-source license, so check the commercial threshold before deploying. Self-hosting requires ~1.4 TB of memory at Q4, so in practice this is an API model — as is every Kimi K2.x release, which are all 1T-class MoE models despite their 32B active-parameter figure.
Kimi K3 2.8T needs about 1691 GB of VRAM at Q4_K_M — quantized weights plus framework overhead, before any KV cache. On Apple Silicon that figure comes out of unified memory.
VRAM and speed by quantization
Quoted against NVIDIA RTX 4090 (24 GB). Weights plus overhead only: this model's architecture is unpublished, so no KV cache is added.
| Quant | Memory | VRAM | Speed (est.) | Fit |
|---|---|---|---|---|
| Q2_K 2.63 bpw | 921.3 GB | — | Too big | |
| Q3_K_M 3.41 bpw | 1194.3 GB | — | Too big | |
| Q4_K_M 4.83 bpw | 1691.3 GB | — | Too big | |
| Q5_K_M 5.67 bpw | 1985.3 GB | — | Too big | |
| Q6_K 6.56 bpw | 2296.8 GB | — | Too big | |
| Q8_0 8.50 bpw | 2975.8 GB | — | Too big | |
| F16 16.00 bpw | 5600.8 GB | — | Too big |
Black marker = usable memory on the NVIDIA RTX 4090 (24 GB). Estimates from the memory-bandwidth roofline on the methodology page. Kimi K3 2.8T VRAM calculator →
Get Kimi K3 2.8T running
Kimi K3 2.8T is not practical on a single consumer GPU: it needs about 1691.3 GB at Q4_K_M, more than any single device in our catalogue holds with headroom. How multi-GPU inference works, or rent a multi-GPU server by the hour:
As an Amazon Associate we earn from qualifying purchases. Cloud GPU links are referral links — we may earn a commission at no extra cost to you.
How to run Kimi K3 2.8T
No local Ollama tag. Ollama lists kimi-k3 only as a cloud-hosted ':cloud' tag, which runs on Ollama's servers, not your machine.
Specifications
Preview — The model is released, but these specs are thin or rest on a single source. Individual fields may be wrong.
Preview — The model is released, but these specs are thin or rest on a single source. Individual fields may be wrong.
- Parameters
- 2.8 Trillion (104B active)
- Context window
- 1,000,000
- Architecture
- Mixture-of-Experts (896 experts, 16 active) + Vision
- Provider
- Moonshot AI
- Licence
- Kimi K3 License (open-weight, not OSI open-source)
- Availability
- Weights downloadable
- Practical deployment
- Private cloud / datacenter cluster
- Specified at
- Q4_K_M
- System RAM
- 2048 GB
- Record updated
- 2026-08-15
- Runs with
- Ollama: not supported source
Weights are downloadable and commercial use is permitted, subject to the licence’s acceptable-use terms.
Quality and use cases
Scores as published by the model’s authors or an independent evaluator — quality, not throughput, and not measured by us.
Independent evaluation — not vendor-claimed and not measured by us
| Benchmark | Score | Provenance |
|---|---|---|
| AA Intelligence Index v4.3.2 (Max) | 44 | independent · Artificial Analysis · snapshot 2026-10-13 |
Can I run Kimi K3 2.8T on my GPU?
Kimi K3 2.8T — frequently asked questions
How much VRAM does Kimi K3 2.8T need?
About 1691 GB at Q4_K_M — quantized weights plus framework overhead, before any KV cache. The cache grows with context length and is added on top; the table above folds it in. Apple Silicon counts unified memory toward the same figure.
Does Kimi K3 2.8T run on an RTX 4090 (24 GB)?
No. Kimi K3 2.8T needs about 1691 GB at Q4_K_M, more than a single RTX 4090's 24 GB. It needs a larger card, several GPUs, or Apple Silicon with enough unified memory — or it runs with part of the weights offloaded to system RAM, which is much slower.
How do I run Kimi K3 2.8T locally?
No local Ollama tag. Ollama lists kimi-k3 only as a cloud-hosted ':cloud' tag, which runs on Ollama's servers, not your machine. Running the published tag would send your prompts to a hosted GPU rather than your own machine.