DE

DeepSeek-V2-Lite

DeepSeekGGUF by mradermacherUpdated 2024-05-31
Download on Hugging Face
Parameters15.7B
Context160K
Layers27
Architecturedeepseek2
Smallest quant6.4 GB
Quick answer

DeepSeek-V2-Lite needs about 13.1 GB of memory at Q4_K_M (a 10.4 GB file) with an 8K context window, so it fits on a 16 GB GPU. This repository has 14 quants, from 6.4 GB (Q2_K) to 16.7 GB (Q8_0). Longer context needs more: at 32K tokens the KV cache alone adds about 9.1 GB.

DeepSeek-V2-Lite: 15.7B parameter MoE model with efficient inference

DeepSeek-V2-Lite is a 15.7B parameter mixture-of-experts language model based on the deepseek2 architecture. It supports a 163,840 token context window and is optimized for efficient inference with Multi-head Latent Attention. The model is available in multiple GGUF quantizations.

Good at

  • Efficient inference with low KV cache
  • Large context support
  • MoE architecture for strong performance

Limits

  • No vision support
  • License not specified
  • Requires 8k context for memory estimates

Best for

  • Local inference on a single 40G GPU
  • Fine-tuning on 8x80G GPUs
  • Research on MoE models

Who it’s for. Researchers and developers needing a large MoE model that can run on a single high‑end GPU. The model is suitable for local deployment and fine‑tuning scenarios where efficient inference is critical.

Quants & memory

QuantFileEst. memoryvs VRAMResultDownload
Q2_K6.4 GB 9.2 GB
Too big Get file →
IQ3_XS7.1 GB 9.9 GB
Too big Get file →
IQ3_S7.5 GB 10.3 GB
Too big Get file →
Q3_K_S7.5 GB 10.3 GB
Too big Get file →
IQ3_M7.6 GB 10.3 GB
Too big Get file →
Q3_K_M8.1 GB 10.9 GB
Too big Get file →
Q3_K_L8.5 GB 11.2 GB
Too big Get file →
IQ4_XS8.6 GB 11.4 GB
Too big Get file →
Q4_K_S9.5 GB 12.3 GB
Too big Get file →
Q4_K_M10.4 GB 13.1 GB
Too big Get file →
Q5_K_S11.1 GB 13.9 GB
Too big Get file →
Q5_K_M11.9 GB 14.6 GB
Too big Get file →
Q6_K14.1 GB 16.8 GB
Too big Get file →
Q8_016.7 GB 19.5 GB
Too big Get file →

For 8GB VRAM, use Q2_K (6.43 GB file, 9.2 GB memory at 8k context). For 16GB VRAM, Q4_K_M (10.36 GB file, 13.1 GB memory) or Q5_K_S (11.14 GB file, 13.9 GB memory). For 24GB VRAM, Q6_K (14.07 GB file, 16.8 GB memory) provides a balance between size and quality.

What GPU do you need for DeepSeek-V2-Lite?

These figures add the file size, an fp16 KV cache at an 8K context window (worked out from this model’s own header) and about 0.5 GB for the runtime. They are estimates, so leave a little headroom, and expect to use less context or a smaller quant if you are right at the limit.

  • 4 GB VRAM: no quant fits at an 8K context.
  • 6 GB VRAM: no quant fits at an 8K context.
  • 8 GB VRAM: no quant fits at an 8K context.
  • 12 GB VRAM: the best quant that fits is IQ4_XS (8.6 GB file, about 11.4 GB in use).
  • 16 GB VRAM: the best quant that fits is Q5_K_M (11.9 GB file, about 14.6 GB in use).
  • 24 GB VRAM: the best quant that fits is Q8_0 (16.7 GB file, about 19.5 GB in use).

Which DeepSeek-V2-Lite quant should you choose?

Q4_K_M is the usual starting point for DeepSeek-V2-Lite: it balances file size and quality well. Here it is a 10.4 GB file.

The Q8_0 file is 16.7 GB, about 1.6 times the size of Q4_K_M. It keeps the weights closer to the original, which only matters if you have the memory to spare.

If memory is tight, the smallest file is Q2_K at 6.4 GB, with the quality trade-offs that come with fewer bits. Use it only when memory is very tight. Expect a clear drop in quality, which is usually worse on small models and easier to live with on large ones.

New to quant names? Read GGUF quantization explained.

Specifications

Architecture from the GGUF header

Architecturedeepseek2
Parameters15.7B (approximate, mixture of experts)
Layers27
Trained context163,840 tokens
KV cache (fp16)270 KB per token

Source

Quantized bymradermacher
FamilyDeepSeek
Downloads (HF)2,088
Last updated2024-05-31

Frequently asked questions about DeepSeek-V2-Lite

How much VRAM does DeepSeek-V2-Lite need?

About 13.1 GB at Q4_K_M with an 8K context window. That is the 10.4 GB file, plus the KV cache and roughly 0.5 GB of runtime overhead. Smaller quants need less and longer context needs more.

Can I run DeepSeek-V2-Lite on an 8 GB GPU?

Not at an 8K context. The smallest file needs about 9.2 GB, so you would need a larger GPU or CPU offloading.

What is the file size of DeepSeek-V2-Lite GGUF?

It depends on the quant: Q2_K is 6.4 GB, IQ3_XS is 7.1 GB, IQ3_S is 7.5 GB, Q3_K_S is 7.5 GB, IQ3_M is 7.6 GB, Q3_K_M is 8.1 GB and 8 more.

What context length does DeepSeek-V2-Lite support?

The GGUF header lists a trained context length of 163,840 tokens (160K). The KV cache grows with the context you use: about 2.3 GB at 8K and 9.1 GB at 32K.

Where can I download DeepSeek-V2-Lite GGUF?

The files are on Hugging Face in the mradermacher/DeepSeek-V2-Lite-GGUF repository. The download buttons on this page open the file on Hugging Face. ggufmodels.com does not host any model files.

Can I run DeepSeek-V2-Lite on 8GB VRAM?

Yes, the Q2_K quantization requires 9.2 GB memory at 8k context, which is close to 8GB VRAM but may need swapping or lower context.

Does the model support vision tasks?

No, DeepSeek-V2-Lite is a text-only model with no vision capabilities.

What is the context length of the model?

The model was trained with a 163,840 token context window.

Similar models

Browse all →

More from mradermacher

All by mradermacher →

More DeepSeek models

All DeepSeek →

Other medium models

All medium models →