Can I run Kimi Linear 48B A3B Instruct locally?
Kimi Linear 48B A3B Instruct by moonshotai has 49B parameters (3.0B active per token, 256 experts, 8 used per token). At 4-bit with an 8K context it needs about 28.9 GB of GPU memory. It fits on 33 of 65 GPUs and Macs we track. See all models that fit in 32 GB of VRAM.
Memory needed by quant and context length
GB of GPU memory = weights + KV cache (FP16) + runtime overhead. Weights use real file sizes from bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF where available (marked •), otherwise parameter count × bits per weight. How we calculate.
| Quant | Quality | 4K ctx | 8K ctx | 32K ctx | 128K ctx |
|---|---|---|---|---|---|
| 2-bit (Q2_K) • | rough | 17.2 GB | 17.3 GB | 18.0 GB | 20.8 GB |
| 3-bit (Q3_K_M) • | usable | 22.0 GB | 22.1 GB | 22.8 GB | 25.6 GB |
| 4-bit (Q4_K_M) • | good, most popular | 28.8 GB | 28.9 GB | 29.7 GB | 32.5 GB |
| 5-bit (Q5_K_M) • | very good | 33.5 GB | 33.6 GB | 34.3 GB | 37.2 GB |
| 6-bit (Q6_K) • | near lossless | 38.5 GB | 38.6 GB | 39.3 GB | 42.2 GB |
| 8-bit (Q8_0) • | effectively lossless | 49.5 GB | 49.6 GB | 50.3 GB | 53.2 GB |
| 16-bit (F16/BF16) • | original weights | 92.4 GB | 92.5 GB | 93.2 GB | 96.0 GB |
Which GPUs and Macs run Kimi Linear 48B A3B Instruct?
Best quality quant that fits at 8K context. "Tight" = within 10% of usable memory. Speed is a ceiling from memory bandwidth (real speed is typically 50–80% of it).
Questions
How much VRAM does Kimi Linear 48B A3B Instruct need?
About 28.9 GB at 4-bit (Q4_K_M) with an 8K context: 28.0 GB of weights (real GGUF file size), 0.2 GB of KV cache and 0.7 GB of runtime overhead. At 8-bit it needs about 49.6 GB.
What is the smallest GPU that runs Kimi Linear 48B A3B Instruct?
No single consumer GPU in our table fits it at Q4_K_M with 8K context; the smallest option that does is the A100 40GB (40 GB).
Can I run Kimi Linear 48B A3B Instruct on a Mac?
Yes. At Q4_K_M with 8K context it fits on a Mac M4 Pro 64GB or larger, using the default macOS GPU memory limit.
Similar models
Same family or a similar parameter count, with the VRAM each needs at Q4_K_M and 8K context.
Kimi Dev 72B
73B · 44.6 GB at 4-bitKimi VL A3B Thinking 2506
16B · 10.8 GB at 4-bitKimi VL A3B Instruct
16B · 10.8 GB at 4-bitNous Hermes 2 Mixtral 8x7B DPO
47B · 28.0 GB at 4-bitfalcon 40b
42B · 25.2 GB at 4-bitflint
40B · 23.6 GB at 4-bitModel facts
| Hugging Face repo | moonshotai/Kimi-Linear-48B-A3B-Instruct |
| Parameters | 49,122,681,728 (49B) · ~3.0B active (from the model name) |
| Layers / hidden size | 27 / 2304 |
| Attention | MLA; 32 heads, 32 KV heads, head dim 72 |
| KV cache per token (FP16) | 30.4 KiB |
| Max context | — |
| License | mit |
| Released on HF | Oct 30, 2025 |
| Downloads (30d) / likes | 48,989 / 618 |
Sources: Hugging Face model API · config.json (moonshotai/Kimi-Linear-48B-A3B-Instruct) · GGUF files · GPU specs from vendor pages (list). Data fetched Nov 17, 2026. Spot an error? Report it.