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 14 GB at 64k, counting the weights, the KV cache and runtime buffers. The cheapest hardware in Canada that runs Kimi VL A3B fully in memory at 64k context is the Intel Arc A770 16GB, $655 at Amazon, at about 37 tok/s (estimated).
Updated . Speeds are estimates; how they are worked out.
What to run it on in Canada
Fully in memory at 64k context, priced at retailers today.
Cheapest
Intel Arc A770 16GB37tok/sest.mlx-5bit
$655
Best value
Intel Arc A770 16GB37tok/sest.mlx-5bit
$655
Fastest
NVIDIA RTX PRO 6000 Blackwell126tok/sest.mlx-8bit
$3,133
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 |
|---|---|---|---|---|---|
| Intel Arc A770 16GB | mlx-5bit | fits | 37 tok/s | $655 | 56.0 |
| AMD Radeon RX 9060 XT 16GB | mlx-5bit | fits | 25 tok/s | $800 | 31.4 |
| AMD Radeon RX 9070 XT | mlx-5bit | fits | 50 tok/s | $1,050 | 47.9 |
| Intel Arc Pro B60 | mlx-8bit | fits | 24 tok/s | $1,060 | 22.4 |
| NVIDIA GeForce RTX 5060 Ti 16GB | mlx-5bit | fits | 40 tok/s | $1,120 | 35.4 |
| AMD Radeon RX 7800 XT | mlx-5bit | fits | 49 tok/s | $1,397 | 35.1 |
| AMD Radeon RX 7600 XT | mlx-5bit | fits | 23 tok/s | $1,440 | 15.7 |
| NVIDIA GeForce RTX 5070 Ti | mlx-5bit | fits | 79 tok/s | $1,690 | 46.9 |
| NVIDIA GeForce RTX 4070 Ti Super | mlx-5bit | fits | 60 tok/s | $1,790 | 33.2 |
| NVIDIA GeForce RTX 4080 Super | mlx-5bit | fits | 65 tok/s | $2,000 | 32.6 |
| AMD Radeon RX 7900 XT | mlx-6bit | fits | 58 tok/s | $2,083 | 27.7 |
| NVIDIA GeForce RTX 5080 | mlx-5bit | fits | 85 tok/s | $2,370 | 35.9 |
| NVIDIA GeForce RTX 3090 | mlx-8bit | fits | 66 tok/s | $2,800 | 23.5 |
| AMD Radeon AI PRO R9700 | mlx-8bit | fits | 40 tok/s | $2,800 | 14.3 |
| NVIDIA RTX PRO 6000 Blackwell | mlx-8bit | fits | 126 tok/s | $3,133 | 40.2 |
| NVIDIA GeForce RTX 3090 Ti | mlx-8bit | fits | 71 tok/s | $3,999 | 17.7 |
| GMKtec EVO-X2 Ryzen AI Max+ 395 128GB | mlx-8bit | fits | 16 tok/s | $5,099 | 3.1 |
| NVIDIA GeForce RTX 4090 | mlx-8bit | fits | 71 tok/s | $7,107 | 10.0 |
| ASUS Ascent GX10 128GB | mlx-8bit | fits | 19 tok/s | $8,299 | 2.3 |
| NVIDIA GeForce RTX 5090 | mlx-8bit | fits | 126 tok/s | $10,300 | 12.2 |
| Mac Studio M4 Max 128GB | mlx-8bit | fits | 40 tok/s | – | – |
| Mac mini M4 Pro 64GB | mlx-8bit | fits | 20 tok/s | – | – |
| AMD Radeon RX 7900 XTX | mlx-8bit | fits | 60 tok/s | – | – |
| Framework Desktop Ryzen AI Max+ 395 128GB | mlx-8bit | fits | 16 tok/s | – | – |
| Mac Studio M3 Ultra 512GB | mlx-8bit | fits | 60 tok/s | – | – |
| NVIDIA DGX Spark 128GB | mlx-8bit | fits | 19 tok/s | – | – |
| Mac Studio M3 Ultra 256GB | mlx-8bit | fits | 60 tok/s | – | – |
| Mac mini M4 Pro 48GB | mlx-8bit | fits | 20 tok/s | – | – |
| NVIDIA GeForce RTX 4060 Ti 16GB | mlx-5bit | fits | 26 tok/s | – | – |
| Mac Studio M4 Max 64GB | mlx-8bit | fits | 40 tok/s | – | – |
| Mac Studio M3 Ultra 96GB | mlx-8bit | fits | 60 tok/s | – | – |
| Framework Desktop Ryzen AI Max+ 395 64GB | mlx-8bit | fits | 16 tok/s | – | – |
| Mac mini M4 32GB | mlx-8bit | fits | 9 tok/s | – | – |
| NVIDIA Tesla P40 | mlx-8bit | fits | 24 tok/s | – | – |
| NVIDIA GeForce RTX 3060 12GB | mlx-5bit | offload | 20 tok/s | $497 | 39.2 |
| Intel Arc B580 | mlx-5bit | offload | 17 tok/s | $560 | 30.9 |
| NVIDIA GeForce RTX 3080 10GB | mlx-5bit | offload | 18 tok/s | $769 | 23.3 |
| NVIDIA GeForce RTX 5070 | mlx-5bit | offload | 25 tok/s | $1,181 | 21.4 |
| NVIDIA GeForce RTX 4070 Super | mlx-5bit | offload | 23 tok/s | $1,400 | 16.2 |
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 14 GB at 64k, counting the weights, the KV cache and runtime buffers.
- What is the best GPU for running Kimi VL A3B locally?
- The cheapest hardware in Canada that runs Kimi VL A3B fully in memory at 64k context is the Intel Arc A770 16GB, $655 at Amazon, at about 37 tok/s (estimated). The fastest we price is the NVIDIA RTX PRO 6000 Blackwell, at about 126 tok/s.
- Can I run Kimi VL A3B on a 16 GB GPU?
- Yes: at mlx-5bit it fits fully in 16 GB at 64k context. On the NVIDIA GeForce RTX 5080 it runs at about 85 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 64k context. On the NVIDIA GeForce RTX 3090 Ti it runs at about 71 tok/s (estimated).
- What licence is Kimi VL A3B released under?
- MIT, which allows commercial use.