nebius/MTP-DeepSeek-V3-0324
Text Generation • 13B • Updated
• 65 • 2
LK-Speculators
High-performance speculative decoding draft models trained using LK losses, a novel training objectives that directly optimize acceptance rate
-
-
nebius/EAGLE3-Llama-3.1-8B-Instruct
Text Generation • 0.4B • Updated • 40 -
nebius/MEDUSA-Llama-3.1-8B-Instruct
Text Generation • 3B • Updated • 33 -
nebius/MLP-Speculator-Llama-3.1-8B-Instruct
Text Generation • 0.2B • Updated • 33 -
nebius/EAGLE3-gpt-oss-20b
Text Generation • 0.4B • Updated • 32 -
nebius/EAGLE3-gpt-oss-120b
Text Generation • 0.4B • Updated • 38 -
nebius/EAGLE3-Llama-3.3-70B-Instruct
Text Generation • 1B • Updated • 27 -
nebius/EAGLE3-Qwen3-235B-A22B-Instruct-2507
Text Generation • 0.6B • Updated • 27 -
LK Losses: Direct Acceptance Rate Optimization for Speculative Decoding
Paper • 2602.23881 • Published • 18