Unverified50% confidenceFactExact time
Meta provides two dedicated draft models for speculative decoding: DFlash V2 is designed for SGLang and is reportedly achieving a 3.7x throughput improvement on AMD MI355X hardware, accepting an average of 6.4 draft tokens per step; DSpark is designed for vLLM and proposes up to seven tokens per pass.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/16/2026
First Seen
Valid until: 1/14/2027
Sources
Related Entities
Related Claims
UnverifiedLlama 3.2 1B Instruct is a lightweight instruction-tuned model released by Meta in 2024 with approximately 1 billion parameters, designed specifically for edge devices.69% similarUnverifiedMeta simultaneously released a draft model called D-Flash, used for speculative decoding to accelerate generation.65% similarUnverifiedV4-Pro uses a Mixture of Experts (MoE) architecture with a total parameter count of 1.6 trillion (1.6T) and 49 billion (49B) active parameters.65% similarUnverifiedMeta's LLaMA 3 series has a maximum scale of 405B, and DeepSeek-V3 is approximately 671B (mixture-of-experts architecture).64% similarUnverifiedLFM2.5 uses a Mixture of Experts (MoE) architecture, activating only approximately 1.5 billion parameters when generating each token.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1047131API
curl https://kongchang.com/api/v1/knowledge/claims/1047131MCP
get_claim(id=1047131)