Unverified90% confidenceFactTime unknown
Step 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
1
Sources
90%
Confidence
Long-term
Relevance
6/1/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedReal-world testing of Qwen 3.6 MTP shows that approximately 20% inference speedup can be achieved through three lines of parameter configuration.65% similarUnverifiedDeveloper Ivan tested version MTP-LX 0.3.5 and achieved a 93.3% accuracy rate on a math benchmark within 5 minutes and 30 seconds.59% similarUnverified基准测试显示tiktoken比HuggingFace的tokenizers库快3-6倍58% similarUnverifiedIn a local inference speed comparison, Agent A1 (95 tokens/second) was faster than Gemma 4 Coder and the Qwen series.58% similarUnverifiedStep 3.7 Flash 在 ClawEval-1.1 基准测试中以 67.1 分领先58% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/21476API
curl https://kongchang.com/api/v1/knowledge/claims/21476MCP
get_claim(id=21476)