Unverified50% confidenceSolutionExact time
Strata uses the model's built-in multi-token prediction layer for speculative decoding, drafting up to 3 tokens at a time and producing an average of 2.4 to 3.2 tokens per pass, with the output results unchanged.
1
Sources
50%
Confidence
Long-term
Relevance
10/5/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn speculative sampling, when the acceptance rate of candidate tokens is high, the effective number of output tokens per forward pass can reach 2–4, thereby improving the overall generation speed by 1.5x–3x.64% similarUnverifiedStep 3.5 Flash内置MTP3（Multi-Token Prediction 3）技术，每次推理同时预测3个token，在编码任务上达到每秒350个token的生成速度63% similarUnverifiedIn speculative decoding, the draft model's parameter count is typically about 1/10 to 1/5 of the target model, and the candidate count k is typically 4-8.63% similarUnverifiedFull-parameter fine-tuning typically requires GPU memory 3–5 times the model's parameter count.61% similarUnverifiedIn the same round of Artificial Analysis evaluations, Qwen 3.8 generated approximately 160 million tokens, while the median for open models in the same tier was 43 million tokens.61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/978619API
curl https://kongchang.com/api/v1/knowledge/claims/978619MCP
get_claim(id=978619)