Kimi-VL-A3B-Thinking-2506
Kimi-VL-A3B-Thinking-2506 is an open-weight vision and language model from Moonshot AI with 16.4B parameters (mixture of experts). At 4-bit it needs about 12.8 GB of memory, so the smallest current device that fits it is the GeForce RTX 5060 Ti 16GB.
Overview
Includes room for an 8k context and runtime overhead. More context needs more memory.
Key information
- Lab
- Moonshot AI
- Released
- 21 Jun 2025
- Parameters
- 16.4B · MoE
- Type
- Vision and language
- Context
- 131,072 tokens
- License
- MIT
- HF downloads, 30 days
- 11,696
- Repository
- moonshotai/Kimi-VL-A3B-Thinking-2506
More from Moonshot AI
View allWill it run on your hardware?
How we estimate| Estimate or device | Memory |
|---|---|
| Q4 | 12.8 GB |
| Q8 | 21.1 GB |
| BF16 | 38.1 GB |
| RTX 5060 Ti 16 GB | 16 GB usable |
| RTX 5090 32 GB | 32 GB usable |
| DGX Spark 128 GB | 108.8 GB usable |
| Mac Studio M5 Ultra 256 GB | 192 GB usable |
Single-device memory estimate. Pick a device to update the On your device column.
36 GB usable memory
| Format | Memory needed | On your device | Example hardware |
|---|---|---|---|
| 4-bit (Q4) | ≈ 12.8 GB | Fits 36% | GeForce RTX 5060 Ti 16GB16 GB memory |
| 8-bit (Q8) | ≈ 21.1 GB | Fits 59% | Arc Pro B6024 GB memory |
| 16-bit (BF16) | ≈ 38.1 GB | Too big 106% | L40S48 GB memory |
Fits means the estimate stays within 90% of usable memory. Tight is above 90% and up to 100%. Too big is above 100%. These are memory checks, not speed or compatibility tests.
Mac allowances are 66% of total RAM up to 36 GB, and 75% above. AI boxes allow 85% because their memory is shared with the operating system.
Where to run it
On a Mac
Start with the MLX files, built for Apple Silicon.
mlx-community/Kimi-VL-A3B-Thinking-2506-6bitCheck your MacWith GGUF
Choose a file for llama.cpp or a compatible app.
ggml-org/Kimi-VL-A3B-Thinking-2506-GGUFFind compatible hardwareThrough an API
No hosted API found.
mlx_vlm.generate --model mlx-community/Kimi-VL-A3B-Thinking-2506-6bit \
--prompt "Describe this image." --image image.jpgllama-cli -hf ggml-org/Kimi-VL-A3B-Thinking-2506-GGUF \
-c 8192 -p "Explain how a transformer works."Example commands built from the repository ids above. They need a working install of MLX or llama.cpp. We have not run them; check each repository for available files.
More about this model
Frequently asked questions
How much memory does Kimi-VL-A3B-Thinking-2506 need? At an 8k context, about 12.8 GB at 4-bit, 21.1 GB at 8-bit and 38.1 GB at 16-bit (BF16). A longer context needs more.
Can I run Kimi-VL-A3B-Thinking-2506 on a Mac? The smallest current Mac that fits the 4-bit estimate of 12.8 GB is the Mac mini M5 Pro (24GB). An MLX build is available: mlx-community/Kimi-VL-A3B-Thinking-2506-6bit.
Does the full context fit? Kimi-VL-A3B-Thinking-2506 supports up to 131,072 tokens of context. Our memory estimates use 8k, so long prompts need more memory than shown.
What does it cost through an API? We did not find a hosted API for this model.
License in plain English
MIT. This license permits commercial use, modification and redistribution, as long as the copyright and license notice are kept. This summary is not legal advice.
OMHQ RigBench
Not yet tested. We show no speed or quality score without a measurement.
Revision history and sources
Last updated . Sources:Hugging Face model card, MLX repository, GGUF repository.