KaLM Embedding Gemma3 12B 2511 vs Qwen 3 Embedding 8B
MTEB multilingual v2 · Mean over tasks · multilingual (1037 languages) Embeddings assign examples to categories, group similar examples, estimate semantic similarity between pairs, classify relationships between pairs, rank relevant documents for queries, find translated sentence pairs, rerank documents according to instructions, assign multiple labels to each example, reorder candidate documents by relevance on MTEB Multilingual v2 · mean over tasks in 1037 languages including 'Auhelawa, Abau, Achi, Achinese, Achuar-Shiwiar, Adzera, Afrikaans, Agarabi; tasks receive equal weight
KaLM Embedding Gemma3 12B 251172.3
Qwen 3 Embedding 8B70.6
| Benchmark | KaLM Embedding Gemma3 12B 2511 | Qwen 3 Embedding 8B |
|---|---|---|
| MTEB multilingual v2 · Mean over tasks · multilingual (1037 languages) | 72.3 | 70.6 |
Specs
| Spec | KaLM Embedding Gemma3 12B 2511 | Qwen 3 Embedding 8B |
|---|---|---|
| Price | — | !$0.01 |
| Context | 33K | 33K |
| Weights | Open | Open |
| Released | Nov 2025 | Jun 2025 |
| Published speed | — | 3.3 × vLLM H100 (Baseten) |
| Dimensions | 3,840 | 4,096 |
| Max input | 131K tokens | 33K tokens |
| Memory (4-bit) | 7.1 GB | 4.5 GB |
| Licence | tencent-kalm-embedding-community | Apache-2.0 |
| Commercial use | Restricted | Allowed |
What breaks when you switch
KaLM Embedding Gemma3 12B 2511 → Qwen 3 Embedding 8B
Re-embed your whole corpus›
Vectors from different models can't be mixed; dimensions change from 3,840 to 4,096.
Max input: 131K tokens → 33K tokens
Qwen 3 Embedding 8B → KaLM Embedding Gemma3 12B 2511
Re-embed your whole corpus›
Vectors from different models can't be mixed; dimensions change from 4,096 to 3,840.