List your company

Language Modeling on Language Modeling (test)

33.19PPL (64)HSA-Ultralong
33.072833.863934.65535.4461Jul 3, 2026

Evaluation results

MethodLinks
HSA-UltralongExtra #Param=12.7%2026.0733.19————26.2621.58.815.844.994.54
Dash-Attention-RoPEExtra #Param=0.006%2026.0733.7————26.7418.67510——
InfLLM v2*Extra #Param=0.006%, Head Dimension=128, Note=Reported for reference2026.0733.71————26.7418.554.9513.6864.24—
Full-Attn RoPEExtra #Param=–2026.0733.92————26.8918.684.9610——
HiLS-Attn-HoPEExtra #Param=0.6%2026.0733.97————26.9118.654.944.344.715.95
Full-Attn HoPEExtra #Param=–2026.0734.15————26.9718.734.956.4210—
NSA-RoPEExtra #Param=0.2%2026.0734.15————27.1118.855.017.6211.7519.14
Full-Attn HoPE (extra MLP)Extra #Param=0.6%2026.0734.28————27.1218.764.955.8510—
SWA-RoPEExtra #Param=–2026.0735.38————27.9419.38.959.018.448.47
Naive-BSA-HoPEExtra #Param=–2026.0736.12————29.0620.454.943.948.67—
Qwen3-4B + EAGLE-2 (Group A)Backbone=Qwen3-4B, Method=EAGLE-22026.01—6.35550.7430——————
Qwen3-4B + FBS-Full (ours)Backbone=Qwen3-4B, Modules=PAW + CH + SG2026.01—6.25320.736——————
Qwen3-4B + FBS-S1Backbone=Qwen3-4B, Modules=PAW + CH2026.01—6.37551.030——————
Qwen3-4B + Lookahead (Group A)Backbone=Qwen3-4B, Method=Lookahead Decoding2026.01—6.45950.8215——————
Qwen3-4B + Medusa (Group A)Backbone=Qwen3-4B, Method=Medusa2026.01—6.55700.818——————
Qwen3-4B + SpecDec (Group A)Backbone=Qwen3-4B, Method=Speculative Decoding2026.01—6.46460.922——————
Qwen3-4B-Instruct (Baseline)Backbone=Qwen3-4B, Variant=Instruct2026.01—6.476010——————