Today for AI
HOT RADAR
computeHEAT 11.5°

AI CURATED HOT STORY · 10/20/2026

Alaya Cloud Releases Alaya-DFlash2, Boosting GLM-5.3 Decoding Speed by 3.41x

1 reports archived1 independent sourcesupdated 10/20/2026, 15:16:04
Synthesis & Latest Updates
1 Sources Cross-Validated

Alaya Cloud released Alaya-DFlash2, a speculative decoding acceleration solution for GLM-5.3 that optimizes long-context generation via parallel draft models and path selection. On 8×H200 GPUs, it achieves 288.2 tokens/s in single-concurrency (3.41x native speed) and improves total throughput across 1-64 concurrency levels. The model is now open-source.

LATEST/Alaya Cloud released Alaya-DFlash2, a speculative decoding acceleration solution for GLM-5.3 that optimizes long-context generation via parallel draft models and path selection. On 8×H200 GPUs, it achieves 288.2 tokens/s in single-concurrency (3.41x native speed) and improves total throughput across 1-64 concurrency levels. The model is now open-source.

HEAT TRENDHourly heat curve

2 fully observed hours
Fewer than 3 fully observed hours; trend not plotted yet.

TIMELINECoverage timeline

Total 1 reports · Latest first
  1. InfoQ 中文T2·68 pts
    • Utilizes parallel draft models and path selection to maintain high acceptance lengths in long contexts
    • Achieves 288.2 tokens/s single-concurrency decoding on 8×H200 (TP=8), 3.41x faster than native autoregressive methods
    • Outperforms native throughput across all tested concurrency levels (1–64), with up to 3x improvement, balancing latency and capacity