Kimi-Linear-48B-A3B-Instruct
(a) On MMLU-Pro (4k context length), Kimi Linear achieves 51.0 performance with similar speed as full attention. On RULER (128k context length), it shows Pareto-optimal performance (84.3) and 3.98x speedup. (b) Kimi Linear achieves 6.3x faster TPOT compared to MLA, offering significant speedups at long sequence lengths (1M tokens).
Params
49.1 B
Context
—
Downloads 30d
191 K
Likes
591
Download history
daily snapshots · 32 days
▲ 6 K in the last 30 days (3.0%)
191 K176 K
Aug 24Sep 3Sep 13Sep 22
191 K176 K
Aug 22Sep 1Sep 12Sep 22
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 | bf16 | 98.2 GB | 115.9 GB | ❌ Won’t fit |
Estimate: file size × 1.1 + KV cache at 8K + 0.5 GB overhead. Not a benchmark — how we calculate this.
Run it
copy-paste, exact tags checked against the Hub$ curl -s https://aimodelscomparison.com/api/v1/models/kimi-linear-48b-a3b-instruct
{
"hf_id": "moonshotai/Kimi-Linear-48B-A3B-Instruct",
"params_b": 49.12,
"context_length": null,
"license": { "id": "mit", "commercial": "yes" },
"downloads_30d": 190802,
"vram_estimates": [
{ "quant": "bf16", "gb": 115.9 }
],
"updated_at": "2026-08-22T01:00:33Z"
}
Specifications
- Architecture
- KimiLinearForCausalLM
- Parameters
- 49.1 B
- Tensor type
- BF16
- Vocabulary
- 163,840
- Layers / heads
- 27 / 32
- Licence
- mit
- First seen on the Hub
- 2025-10-30
- Training datasets
- undisclosed
- Added to our catalog
- 2026-08-22
Family
Base model and the most-downloaded derivatives in the catalog.
Compare with any text-generation model