Kimi-K3
Kimi K3 is an open-weight, native multimodal agentic model and our most capable model to date. It is a 2.8T-parameter model built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), with native vision capabilities and a 1-million-token context window. It is the world's first open 3T-class model, designed for frontier intelligence across long-horizon coding, knowledge work, and reasoning.
Params
2,779.9 B
Context
—
Downloads 30d
856 K
Likes
11,836
Download history
daily snapshots · 99 days
▲ 304 K in the last 30 days (26.2%)
1.1 M856 K
Oct 12Oct 22Nov 1Nov 10
2.9 M856 K
Aug 13Sep 12Oct 12Nov 10
2.9 M856 K
Aug 4Sep 6Oct 9Nov 10
Can you run it?
Estimated VRAM at 8K context unless noted. Pick your hardware to see the verdict per quantization.
| File | Quant | Size | Est. VRAM | Verdict on RTX 4090 · 24 GB |
|---|---|---|---|---|
| model.safetensors | u8 | 1,560.9 GB | 2,134.5 GB | ❌ Won’t fit |
Estimate: file size × 1.1 + KV cache at 8K + 0.5 GB overhead. Not a benchmark — how we calculate this.
Use via API
via OpenRouter · USD per 1M tokens · synced 2026-11-10$0.19
input / 1M
$16.00
output / 1M
$0.18
cached input / 1M
$16.19
1M in + 1M out
:batch variant
$2.28 in · $11.40 out
- Context window
- 1,048,576 tokens
- Max output
- 943,718 tokens
- Providers
- 20
- Cheapest provider
- Wafer · $0.28 / $9.60
- Best throughput
- —
- AA Intelligence / Coding / Agentic
- 17.4 / 72 / 29.3
✓ Tool calling
✓ Reasoning
✓ Structured outputs
✓ Image input
✗ File input
✗ Audio input
✓ Prompt caching
Specifications
- Architecture
- KimiK3ForConditionalGeneration
- Parameters
- 2,779.9 B
- Tensor type
- U8
- Licence
- other
- First seen on the Hub
- 2026-06-13
- Training datasets
- undisclosed
- Added to our catalog
- 2026-08-04
Compare with any image-text-to-text model