DeepSeek-V2-Lite
16B MoE that runs like a 2.4B model, with a 32K context window.
About
DeepSeek-V2-Lite (May 16, 2024) is the small sibling of DeepSeek-V2: 16B total parameters with only 2.4B active per token, trained from scratch on 5.7T tokens. It introduced Multi-head Latent Attention (MLA), which compresses the KV cache into a latent vector for efficient inference. Team-reported: 58.3 MMLU and 41.1 GSM8K for the base model, well ahead of the older 7B dense and 16B MoE baselines, and it deploys on a single 40GB GPU. It ships under DeepSeek's model license, which the model card describes as supporting commercial use — confirm the current terms on Hugging Face. Honest caveats: it is a 2024 model and newer small models have passed it; and as an MoE it still needs the full 16B in memory despite the 2.4B active count.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Team-reported: 58.3 MMLU, 41.1 GSM8K, 60.3 C-Eval (base model).
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (verified tier, metrics-v1). Evidence: 2024-05-16 release verified via Hugging Face model card news section.; 16B total / 2.4B active MoE, MLA architecture, 32K context verified via model card.; Team-reported 58.3 MMLU / 41.1 GSM8K labeled as team-reported.; DeepSeek model license commercial-use claim flagged for confirmation; 2024-vintage caveat included..