Giga-Embeddings: Sparse 10B MoE encoder boosts throughput by 2.65x
_reachsumit · x · 2026-08-26
Egor Kolodin et al. introduced the Giga-Embeddings family, designed to combine strong retrieval quality with efficient serving. The flagship is a sparse 10B-parameter Mixture-of-Experts (MoE) encoder with approximately 1.8B active parameters per token. It achieves top aggregate performance on MTEB benchmarks across English, Russian, multilingual, and code tasks. In vLLM benchmarks with 1024-token inputs, it processes 114.5k tokens per second, offering 25% higher throughput than a dense 3B model and 1.56-2.65x the throughput of evaluated external systems. The family also includes a dense 3B encoder and a distilled 480M encoder, which scores 70.98 on Russian MTEB, surpassing FRIDA with 42% fewer parameters. All checkpoints are released.
More from Models
- How LLMs Self-Correct Mid-Generation: The Role of Reasoning RL and Instructions — dejanseo · 2026-08-26
- User observes A/B responses containing contradictory facts from same sources — Brilliant_Agent_1427 · 2026-08-26
- Discussion on model trends: Fable 5 performance sparks debate — voooooogel · 2026-08-26
- Muse Glimmer outperforms Qwen 3.8 xhigh in benchmarks — Ok-Inevitable8391 · 2026-08-26
- Tests compare upcoming Claude models Marshmallow and Melon — toptickcrypto · 2026-08-26
- Developer reports severe Claude degradation: outputs gibberish and refuses instructions — sheketsilencio · 2026-08-26