Deepseek Vl2 Tiny — Hardware Requirements & GPU Compatibility
Vision
DeepSeek-VL2-Tiny is the smallest in DeepSeek's VL2 series of Mixture-of-Experts vision-language models, with about 3.4 billion total parameters but only roughly 1.2 billion activated per token. It handles vision-language tasks — visual question answering, OCR, document/table/chart understanding, and visual grounding — below the larger VL2-Small and full VL2 variants. Only a fraction of parameters activate per token, so decoding stays fast even though every expert must still be loaded into memory. It is compact enough to run on a single consumer GPU once quantized. Context length is limited to 4,096 tokens. It is released under DeepSeek's own model license, a custom permissive license with an acceptable-use policy rather than a fully open license like MIT or Apache 2.0. Published in December 2024, it pairs a small DeepSeekMoE-3B backbone with a dynamic image-tiling vision encoder.
Runs great on a 24 GB GPU
Runs from 8 GB (BF16 · 4K)Best on 24 GB (BF16 · ~152 tok/s · 32K)Macs from 16 GB (BF16 · ~36 tok/s)
Other
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run Deepseek Vl2 Tiny
One recommended way. Every other method, OS and quant is in the step-by-step guide.Each way opens a step-by-step guide for your OS and quant.
Recommended
vllm serve deepseek-ai/deepseek-vl2-tiny --max-model-len 16384vLLM repo with its own quantisation
Download: Weights on Hugging Face
Ways to runOther ways
each opens step-by-stepRun in cloud
Runs on RTX 3060 12GB (2 GB headroom) · BF16
- Generation speed
- ~73 tok/s
- generation speed
- Cost per 1M output tokens
- $0.23
- per 1M output tokens
How Much VRAM Does Deepseek Vl2 Tiny Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | On RTX 4090 (reference) | |
|---|---|---|---|---|
| 7.4 GB | 11.7 GB | 16-bit · bfloat — Brain floating point 16 — preferred for training | Fits: 32K context · ~152 tok/s |
Selected BF16
Brain floating point 169.4GBat 16K context
- Weights
- 6.8 GB
- KV cache
- 2.3 GB
- Overhead
- 0.3 GB
Fits RTX 4090 (reference) · 14.6 GB free · ~152 tok/s
≥50 tok/s needs ≥229 GB/s · RTX 5090 ~206, RTX 3090 Ti ~152
Which Hardware Can Run Deepseek Vl2 Tiny?
BF16 · 7.4 GBDeepseek Vl2 Tiny (BF16) requires 7.4 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 10+ GB is recommended. 52 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti, NVIDIA GeForce RTX 3080. 151 devices with unified memory can run Deepseek Vl2 Tiny, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), NVIDIA Jetson Orin Nano 8GB (Super).
12 GB
12 GB tier →fits · 2.6 GB free · up to 32K ctx
10 GPUs
RTX 3080 Ti~143Simulate output on RTX 3080 TiRTX 5070~117Simulate output on RTX 5070RTX 4070~95Simulate output on RTX 407020 GB
fits · 10.6 GB free · up to 32K ctx
1 GPU
RX 7900 XT~67Simulate output on RX 7900 XT32–40 GB
fits · 22.6 GB free · up to 32K ctx
1 GPU · 3 Mac classes · 4 device families
RTX 5090~206Simulate output on RTX 5090M5 Max36 GB~57Simulate output on Mac Studio M5 Max (36 GB)M532 GB~30Simulate output on MacBook Air 13" M5 (32 GB)Orin~45Simulate output on NVIDIA Jetson AGX Orin 32GBSnapdragon X Elite (X1E-80-100)~29Simulate output on Snapdragon X Elite Copilot+ PC
By VRAM tier · typical GPUs and Macs at BF16
Alternatives to Deepseek Vl2 Tiny
DeepSeek family · 5 models →Similar models
6| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| D1 3BLiquid AI | — | 2.6 GB | ~285 | same size |
| Granite 4.0 MicroIBM | — | 3.7 GB | ~261 | same size |
| Llama 3.2 3B InstructMeta | — | 4.1 GB | ~309 | same size |
| Tiny Aya GlobalCohere | — | 9.4 GB | ~89 | same size |
| AI21 Jamba Reasoning 3BAI21 Labs | — | 2.3 GB | ~291 | same size |
| Sarashina2.2 3B Instruct v0.1sbintuitions | — | 3.7 GB | ~247 | same size |
Specifications
- Publisher
- DeepSeek
- Family
- DeepSeek
- Parameters
- 3.4B
- Active parameters
- 1.2B per token
- Release Date
- 2024-12-13
- License
- Other
Frequently Asked Questions
How much VRAM does Deepseek Vl2 Tiny need?
Deepseek Vl2 Tiny requires 7.4 GB of VRAM at BF16.
VRAM = Weights + KV Cache + Overhead
Weights = 3.4B × 16 bits ÷ 8 = 6.7 GB
KV Cache + Overhead ≈ 0.7 GB (at 2K context + ~0.3 GB framework)
Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.
VRAM usage by quantization
Can I run Deepseek Vl2 Tiny on a Mac?
Yes — Mac Mini M2 Pro (16 GB) and 133 other Macs can run Deepseek Vl2 Tiny. Apple Silicon uses unified memory, so the model shares RAM with the system. At BF16 you need at least 7.4 GB of usable unified memory (RAM minus macOS overhead).
Can I run Deepseek Vl2 Tiny locally?
Yes — Deepseek Vl2 Tiny can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is Deepseek Vl2 Tiny?
At BF16, Deepseek Vl2 Tiny can reach ~99 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~152 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.
tok/s = 1000 ÷ (active GB ÷ (bandwidth GB/s × efficiency) × 1000 + layers × routing ms)
Mixture-of-Experts: just the active experts get read for each token, along with a fixed routing cost per layer.
Example: NVIDIA B200 → 2.6 GB active ÷ (8000 × 0.65) = 0.498 ms, plus 48 layers × 0.055 ms = 2.640 ms, so 1000 ÷ 3.138 ms = ~319 tok/s
Estimated speed at BF16 (7.4 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of Deepseek Vl2 Tiny?
At BF16, the download is about 6.74 GB.
Which GPUs can run Deepseek Vl2 Tiny?
52 consumer GPUs can run Deepseek Vl2 Tiny at BF16. Top options include AMD Radeon RX 6700 XT, AMD Radeon RX 6800, AMD Radeon RX 6800 XT, AMD Radeon RX 7600. 36 GPUs have plenty of headroom for comfortable inference.
Which devices can run Deepseek Vl2 Tiny?
151 devices with unified memory can run Deepseek Vl2 Tiny at BF16 (7.4 GB), including AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.