Can I run Kimi VL A3B Thinking locally?
Kimi VL A3B Thinking by moonshotai has 16B parameters (3.0B active per token, 64 experts, 6 used per token). At 4-bit with an 8K context it needs about 10.3 GB of GPU memory. It fits on 65 of 65 GPUs and Macs we track. See all models that fit in 12 GB of VRAM.
Memory needed by quant and context length
GB of GPU memory = weights + KV cache (FP16) + runtime overhead. Weights use parameter count × bits per weight. How we calculate.
| Quant | Quality | 4K ctx | 8K ctx | 32K ctx | 128K ctx |
|---|---|---|---|---|---|
| 2-bit (Q2_K) | rough | 6.6 GB | 6.7 GB | 7.4 GB | 10.2 GB |
| 3-bit (Q3_K_M) | usable | 8.3 GB | 8.4 GB | 9.1 GB | 12.0 GB |
| 4-bit (Q4_K_M) | good, most popular | 10.2 GB | 10.3 GB | 11.0 GB | 13.8 GB |
| 5-bit (Q5_K_M) | very good | 11.7 GB | 11.8 GB | 12.5 GB | 15.4 GB |
| 6-bit (Q6_K) | near lossless | 13.4 GB | 13.5 GB | 14.2 GB | 17.0 GB |
| 8-bit (Q8_0) | effectively lossless | 17.1 GB | 17.2 GB | 17.9 GB | 20.7 GB |
| 16-bit (F16/BF16) | original weights | 31.4 GB | 31.5 GB | 32.2 GB | 35.1 GB |
Which GPUs and Macs run Kimi VL A3B Thinking?
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).
| Hardware | Memory | Fits? | Best quant | Needs | Max tok/s (est.) | |
|---|---|---|---|---|---|---|
| RTX 3060 12GB | 12 GB | Fits | Q3_K_M | 8.4 GB | 246 | details |
| RTX 4060 Ti 16GB | 16 GB | Fits | Q6_K | 13.5 GB | 117 | details |
| RTX 3090 | 24 GB | Fits | Q8_0 | 17.2 GB | 294 | details |
| RTX 4090 | 24 GB | Fits | Q8_0 | 17.2 GB | 316 | details |
| RTX 5090 | 32 GB | Fits | Q8_0 | 17.2 GB | 562 | details |
| RTX 5070 Ti | 16 GB | Fits | Q6_K | 13.5 GB | 364 | details |
| RX 7900 XTX | 24 GB | Fits | Q8_0 | 17.2 GB | 301 | details |
| Mac M4 Pro 48GB | 48 GB | Fits | Q8_0 | 17.2 GB | 86 | details |
| Mac M4 Max 128GB | 128 GB | Fits | F16 | 31.5 GB | 91 | details |
| Mac M3 Ultra 512GB | 512 GB | Fits | F16 | 31.5 GB | 137 | details |
| DGX Spark | 128 GB | Fits | F16 | 31.5 GB | 46 | details |
| RTX 4070 | 12 GB | Fits | Q3_K_M | 8.4 GB | 344 | details |
| RTX 5060 Ti 16GB | 16 GB | Fits | Q6_K | 13.5 GB | 182 | details |
| RTX 4060 | 8 GB | Fits | Q2_K | 6.7 GB | 242 | details |
| Arc B580 | 12 GB | Fits | Q3_K_M | 8.4 GB | 311 | details |
| RTX 3060 Ti | 8 GB | Fits | Q2_K | 6.7 GB | 398 | details |
| RTX 3070 | 8 GB | Fits | Q2_K | 6.7 GB | 398 | details |
| RTX 3080 10GB | 10 GB | Fits | Q3_K_M | 8.4 GB | 518 | details |
| RTX 3080 12GB | 12 GB | Fits | Q3_K_M | 8.4 GB | 622 | details |
| RTX 3090 Ti | 24 GB | Fits | Q8_0 | 17.2 GB | 316 | details |
| RTX 4060 Ti 8GB | 8 GB | Fits | Q2_K | 6.7 GB | 256 | details |
| RTX 4070 Super | 12 GB | Fits | Q3_K_M | 8.4 GB | 344 | details |
| RTX 4070 Ti | 12 GB | Fits | Q3_K_M | 8.4 GB | 344 | details |
| RTX 4070 Ti Super | 16 GB | Fits | Q6_K | 13.5 GB | 273 | details |
| RTX 4080 | 16 GB | Fits | Q6_K | 13.5 GB | 291 | details |
| RTX 4080 Super | 16 GB | Fits | Q6_K | 13.5 GB | 299 | details |
| RTX 5060 | 8 GB | Fits | Q2_K | 6.7 GB | 398 | details |
| RTX 5070 | 12 GB | Fits | Q3_K_M | 8.4 GB | 458 | details |
| RTX 5080 | 16 GB | Fits | Q6_K | 13.5 GB | 390 | details |
| 2x RTX 3090 | 2×24 GB | Fits | F16 | 31.5 GB | 156 | details |
| 2x RTX 4090 | 2×24 GB | Fits | F16 | 31.5 GB | 168 | details |
| 2x RTX 5090 | 2×32 GB | Fits | F16 | 31.5 GB | 299 | details |
| RTX A6000 | 48 GB | Fits | F16 | 31.5 GB | 128 | details |
| RTX 6000 Ada | 48 GB | Fits | F16 | 31.5 GB | 160 | details |
| RTX PRO 6000 Blackwell | 96 GB | Fits | F16 | 31.5 GB | 299 | details |
| L40S | 48 GB | Fits | F16 | 31.5 GB | 144 | details |
| A100 40GB | 40 GB | Fits | F16 | 31.5 GB | 259 | details |
| A100 80GB | 80 GB | Fits | F16 | 31.5 GB | 340 | details |
| H100 PCIe | 80 GB | Fits | F16 | 31.5 GB | 333 | details |
| H100 SXM | 80 GB | Fits | F16 | 31.5 GB | 558 | details |
| H200 | 141 GB | Fits | F16 | 31.5 GB | 800 | details |
| RX 7600 | 8 GB | Fits | Q2_K | 6.7 GB | 256 | details |
| RX 7700 XT | 12 GB | Fits | Q3_K_M | 8.4 GB | 295 | details |
| RX 7800 XT | 16 GB | Fits | Q6_K | 13.5 GB | 254 | details |
| RX 7900 GRE | 16 GB | Fits | Q6_K | 13.5 GB | 234 | details |
| RX 7900 XT | 20 GB | Fits | Q6_K | 13.5 GB | 325 | details |
| RX 9070 | 16 GB | Fits | Q6_K | 13.5 GB | 260 | details |
| RX 9070 XT | 16 GB | Fits | Q6_K | 13.5 GB | 260 | details |
| Ryzen AI Max+ 395 128GB | 128 GB | Fits | F16 | 31.5 GB | 43 | details |
| Arc A770 16GB | 16 GB | Fits | Q6_K | 13.5 GB | 228 | details |
| Mac M1 16GB | 16 GB | Fits | Q3_K_M | 8.4 GB | 46 | details |
| Mac M2 24GB | 24 GB | Fits | Q6_K | 13.5 GB | 41 | details |
| Mac M3 24GB | 24 GB | Fits | Q6_K | 13.5 GB | 41 | details |
| Mac M4 16GB | 16 GB | Fits | Q3_K_M | 8.4 GB | 82 | details |
| Mac M4 32GB | 32 GB | Fits | Q8_0 | 17.2 GB | 38 | details |
| Mac M1 Pro 32GB | 32 GB | Fits | Q8_0 | 17.2 GB | 63 | details |
| Mac M3 Pro 36GB | 36 GB | Fits | Q8_0 | 17.2 GB | 47 | details |
| Mac M4 Pro 24GB | 24 GB | Fits | Q6_K | 13.5 GB | 111 | details |
| Mac M4 Pro 64GB | 64 GB | Fits | F16 | 31.5 GB | 46 | details |
| Mac M1 Max 64GB | 64 GB | Fits | F16 | 31.5 GB | 67 | details |
| Mac M2 Max 96GB | 96 GB | Fits | F16 | 31.5 GB | 67 | details |
| Mac M3 Max 128GB | 128 GB | Fits | F16 | 31.5 GB | 67 | details |
| Mac M4 Max 64GB | 64 GB | Fits | F16 | 31.5 GB | 91 | details |
| Mac M2 Ultra 192GB | 192 GB | Fits | F16 | 31.5 GB | 133 | details |
| Mac M3 Ultra 256GB | 256 GB | Fits | F16 | 31.5 GB | 137 | details |
Questions
How much VRAM does Kimi VL A3B Thinking need?
About 10.3 GB at 4-bit (Q4_K_M) with an 8K context: 9.3 GB of weights (estimated), 0.2 GB of KV cache and 0.7 GB of runtime overhead. At 8-bit it needs about 17.2 GB.
What is the smallest GPU that runs Kimi VL A3B Thinking?
At Q4_K_M and 8K context, the smallest single consumer GPU in our table that fits it is the RTX 4060 Ti 16GB (16 GB).
Can I run Kimi VL A3B Thinking on a Mac?
Yes. At Q4_K_M with 8K context it fits on a Mac M2 24GB 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 VL A3B Thinking 2506
16B · 10.8 GB at 4-bitKimi VL A3B Instruct
16B · 10.8 GB at 4-bitMoonlight 16B A3B Instruct
16B · 10.8 GB at 4-bitdeepseek moe 16b base
16B · 12.5 GB at 4-bitLLaDA2.1 mini
16B · 10.3 GB at 4-bitDeepSeek Coder V2 Lite Instruct
16B · 10.5 GB at 4-bitModel facts
| Hugging Face repo | moonshotai/Kimi-VL-A3B-Thinking |
| Parameters | 16,407,657,776 (16B) · ~3.0B active (from the model name) |
| Layers / hidden size | 27 / 2048 |
| Attention | MLA; 16 heads, 16 KV heads, head dim 128 |
| KV cache per token (FP16) | 30.4 KiB |
| Max context | 131,072 tokens |
| License | mit |
| Released on HF | Apr 9, 2025 |
| Downloads (30d) / likes | 8,510 / 459 |
Sources: Hugging Face model API · config.json (moonshotai/Kimi-VL-A3B-Thinking) · GPU specs from vendor pages (list). Data fetched Nov 17, 2026. Spot an error? Report it.