All tasks / Text & knowledge / Embeddings & retrieval
Best local model for embeddings & retrieval on a 48 GB GPU
Open-weights models whose approximate minimum GPU memory is 48 GB or less, ranked from public benchmarks. Memory figures are approximate; quantised builds can need less.
STop ranked: KaLM-Embedding-Gemma3-12B-2511 (Tencent), 72.3% on MTEB(Multilingual, v2).9 options listed · data verified 19 Oct 2026
Ranked list
| Tier | Name | Approach | Runs | Rated by | Price | Skill | Compare |
|---|---|---|---|---|---|---|---|
| S |  KaLM-Embedding-Gemma3-12B-2511Tencent | Models | Local | 72.3% on MTEB(Multilingual, v2) | ~24 GB VRAM | — | |
| S |  Qwen3-Embedding-8BAlibaba (Qwen) | Models | Local | 70.6% on MTEB(Multilingual, v2) | ~16 GB VRAM | — | vs KaLM-Embedding-Gemma3-12B-2511 |
| A |  harrier-oss-v1-0.6bMicrosoft | Models | Local | 69% on MTEB(Multilingual, v2) | ~2 GB VRAM | — | vs KaLM-Embedding-Gemma3-12B-2511 |
| A |  F2LLM-v2-14BAnt Group (codefuse-ai) | Models | Local | 68.7% on MTEB(Multilingual, v2) | ~28 GB VRAM | — | vs KaLM-Embedding-Gemma3-12B-2511 |
| A |  jina-embeddings-v5 (omni-small)Jina AI (Elastic) | Models | Local | 67% on MTEB(Multilingual, v2) | ~4 GB VRAM | — | |
| B |  nomic-embed-text-v2-moeNomic AI | Models | Local | 57.6% on MTEB(Multilingual, v2) | ~2 GB VRAM | — | |
| B |  snowflake-arctic-embed-l-v2.0Snowflake | Models | Local | 57% on MTEB(Multilingual, v2) | ~3 GB VRAM | — | |
| B |  zembed-1ZeroEntropy | Models | Local | 56.6% on MTEB(Multilingual, v2) | ~8 GB VRAM | — | |
| C |  granite-embedding-311m-multilingual-r2IBM | Models | Local | 56% on MTEB(Multilingual, v2) | ~1 GB VRAM | — |
Tiers come only from public benchmarks (MTEB(Multilingual, v2)); tools without a result of their own inherit the tier of the best model they offer. How we rank.