Kimi VL A3B
16.4B total, 3B active · 128k context · MIT · released 2025-04-09
visionagenticchatreasoning
In short
Kimi VL A3B is an open model from Moonshot AI with 16.4B parameters (3B active per token). At mlx-5bit, Kimi VL A3B needs about 12 GB of memory at 8k context and 12 GB at 8k, counting the weights, the KV cache and runtime buffers. The cheapest hardware in the US that runs Kimi VL A3B fully in memory at 8k context is the NVIDIA Tesla P40, $345 at Amazon, at about 37 tok/s (estimated).
Updated . Speeds are estimates; how they are worked out.
What to run it on in the US
Fully in memory at 8k context, priced at retailers today.
Cheapest
NVIDIA Tesla P4037tok/sest.mlx-8bit
$345
Best value
Intel Arc A770 16GB55tok/sest.mlx-6bit
$400
Fastest
NVIDIA GeForce RTX 5090190tok/sest.mlx-8bit
$4,921
Mac
Mac mini M4 Pro 48GB30tok/sest.mlx-8bit
$2,149
All hardware that runs it
Everything we price, best value first. Two cards split the model by layers.
| Hardware | Quant | Fits | Decode | Price | tok/s per 1,000 |
|---|---|---|---|---|---|
| NVIDIA Tesla P40 | mlx-8bit | fits | 37 tok/s | $345 | 106.1 |
| Intel Arc A770 16GB | mlx-6bit | fits | 55 tok/s | $400 | 138.3 |
| AMD Radeon RX 7600 XT | mlx-6bit | fits | 34 tok/s | $480 | 71.0 |
| AMD Radeon RX 9060 XT 16GB | mlx-6bit | fits | 38 tok/s | $520 | 72.9 |
| AMD Radeon RX 7800 XT | mlx-6bit | fits | 74 tok/s | $550 | 134.5 |
| NVIDIA GeForce RTX 5060 Ti 16GB | mlx-6bit | fits | 60 tok/s | $618 | 96.9 |
| Intel Arc Pro B60 | mlx-8bit | fits | 36 tok/s | $650 | 54.9 |
| NVIDIA GeForce RTX 4060 Ti 16GB | mlx-6bit | fits | 39 tok/s | $770 | 50.0 |
| AMD Radeon RX 7900 XT | mlx-8bit | fits | 75 tok/s | $800 | 93.9 |
| AMD Radeon RX 9070 XT | mlx-6bit | fits | 76 tok/s | $815 | 93.1 |
| NVIDIA GeForce RTX 4070 Ti Super | mlx-6bit | fits | 90 tok/s | $984 | 91.3 |
| NVIDIA GeForce RTX 5070 Ti | mlx-6bit | fits | 120 tok/s | $995 | 120.3 |
| NVIDIA GeForce RTX 4080 Super | mlx-6bit | fits | 98 tok/s | $1,199 | 82.1 |
| Mac mini M4 32GB | mlx-8bit | fits | 13 tok/s | $1,420 | 9.4 |
| AMD Radeon RX 7900 XTX | mlx-8bit | fits | 90 tok/s | $1,533 | 58.8 |
| NVIDIA GeForce RTX 3090 | mlx-8bit | fits | 99 tok/s | $1,600 | 61.9 |
| NVIDIA GeForce RTX 5080 | mlx-6bit | fits | 128 tok/s | $1,600 | 80.2 |
| AMD Radeon AI PRO R9700 | mlx-8bit | fits | 60 tok/s | $1,700 | 35.4 |
| NVIDIA GeForce RTX 3090 Ti | mlx-8bit | fits | 107 tok/s | $2,060 | 51.8 |
| NVIDIA GeForce RTX 4090 | mlx-8bit | fits | 107 tok/s | $2,073 | 51.5 |
| Mac mini M4 Pro 48GB | mlx-8bit | fits | 30 tok/s | $2,149 | 14.1 |
| Mac mini M4 Pro 64GB | mlx-8bit | fits | 30 tok/s | $2,900 | 10.4 |
| Mac Studio M4 Max 64GB | mlx-8bit | fits | 61 tok/s | $3,400 | 17.8 |
| Framework Desktop Ryzen AI Max+ 395 128GB | mlx-8bit | fits | 24 tok/s | $3,449 | 7.0 |
| GMKtec EVO-X2 Ryzen AI Max+ 395 128GB | mlx-8bit | fits | 24 tok/s | $3,650 | 6.6 |
| Mac Studio M3 Ultra 96GB | mlx-8bit | fits | 91 tok/s | $4,399 | 20.7 |
| NVIDIA GeForce RTX 5090 | mlx-8bit | fits | 190 tok/s | $4,921 | 38.5 |
| Mac Studio M4 Max 128GB | mlx-8bit | fits | 61 tok/s | $4,999 | 12.1 |
| ASUS Ascent GX10 128GB | mlx-8bit | fits | 29 tok/s | $6,100 | 4.7 |
| NVIDIA DGX Spark 128GB | mlx-8bit | fits | 29 tok/s | $7,399 | 3.9 |
| NVIDIA RTX PRO 6000 Blackwell | mlx-8bit | fits | 190 tok/s | $7,997 | 23.7 |
| Mac Studio M3 Ultra 256GB | mlx-8bit | fits | 91 tok/s | $8,499 | 10.7 |
| Mac Studio M3 Ultra 512GB | mlx-8bit | fits | 91 tok/s | $23,000 | 4.0 |
| Framework Desktop Ryzen AI Max+ 395 64GB | mlx-8bit | fits | 24 tok/s | – | – |
| Intel Arc B580 | mlx-5bit | offload | 41 tok/s | $346 | 118.8 |
| NVIDIA GeForce RTX 3060 12GB | mlx-5bit | offload | 47 tok/s | $360 | 131.7 |
| NVIDIA GeForce RTX 3080 10GB | mlx-5bit | offload | 42 tok/s | $469 | 89.1 |
| NVIDIA GeForce RTX 4070 Super | mlx-5bit | offload | 62 tok/s | $699 | 88.1 |
| NVIDIA GeForce RTX 5070 | mlx-5bit | offload | 76 tok/s | $750 | 101.2 |
1 other card or machine we price cannot hold it at this context.
Memory needed
Weights, KV cache and runtime buffers at each context length (KV cache at fp16).
| Quant | File | @8k | @32k | @64k | @128k |
|---|---|---|---|---|---|
| mlx-5bitmlx | 11.0 GB | 12.0 GB | 12.9 GB | 14.0 GB | 16.3 GB |
| mlx-6bitmlx | 12.9 GB | 13.9 GB | 14.7 GB | 15.9 GB | 18.1 GB |
| mlx-8bitmlx | 16.6 GB | 17.6 GB | 18.4 GB | 19.6 GB | 21.8 GB |
Run it
Copy a command for the runtime you use.
- mlx-5bit
- LM Studio
lms get mlx-community/Kimi-VL-A3B-Thinking-2506-5bit - MLX
mlx_lm.server --model mlx-community/Kimi-VL-A3B-Thinking-2506-5bit
- LM Studio
- mlx-6bit
- LM Studio
lms get mlx-community/Kimi-VL-A3B-Thinking-2506-6bit - MLX
mlx_lm.server --model mlx-community/Kimi-VL-A3B-Thinking-2506-6bit
- LM Studio
- mlx-8bit
- LM Studio
lms get mlx-community/Kimi-VL-A3B-Thinking-2506-8bit - MLX
mlx_lm.server --model mlx-community/Kimi-VL-A3B-Thinking-2506-8bit
- LM Studio
Notes
MoE vision-language model with 16B total and about 3B active parameters; it needs trust_remote_code and was developed against transformers 4.48.2. vLLM support needed a branch from an unmerged pull request (#16387) when the card was written.
Questions
- How much VRAM does Kimi VL A3B need?
- At mlx-5bit, Kimi VL A3B needs about 12 GB of memory at 8k context and 12 GB at 8k, counting the weights, the KV cache and runtime buffers.
- What is the best GPU for running Kimi VL A3B locally?
- The cheapest hardware in the US that runs Kimi VL A3B fully in memory at 8k context is the NVIDIA Tesla P40, $345 at Amazon, at about 37 tok/s (estimated). The best value is the Intel Arc A770 16GB (about 55 tok/s for $400). The fastest we price is the NVIDIA GeForce RTX 5090, at about 190 tok/s.
- Can I run Kimi VL A3B on a 16 GB GPU?
- Yes: at mlx-6bit it fits fully in 16 GB at 8k context. On the NVIDIA GeForce RTX 5080 it runs at about 128 tok/s (estimated).
- Can I run Kimi VL A3B on a 24 GB GPU?
- Yes: at mlx-8bit it fits fully in 24 GB at 8k context. On the NVIDIA GeForce RTX 3090 Ti it runs at about 107 tok/s (estimated).
- What licence is Kimi VL A3B released under?
- MIT, which allows commercial use.