Qwen3-1.7B-Base — all DFlash lab methods

Winner: Normal Qwen KV-cache at 4.114 tok/s (1.000× normal cached Qwen).

Frozen target: Qwen/Qwen3-1.7B-Base · 1,720,574,976 parameters · float32 CPU · decode-only timing.

Training corpus: 3,040 sliding blocks, 9.74× the earlier Qwen3-0.6B experiment. Candidate coverage: 98.9%.

Method	tok/s	vs normal	draft acceptance	target calls	target input tokens	tokens / target call	guidance scores	exact
Normal Qwen KV-cache 🏆	4.114	1.000×	0.0%	15.00	15.00	1.07	0	✓
DFlash hidden fusion	1.610	0.391×	5.8%	12.67	67.67	1.29	0	✓
DFlash2 DP selector	1.782	0.433×	8.0%	11.83	63.50	1.40	2644	✓
DFlash3-MOBS	1.776	0.432×	8.1%	11.83	63.33	1.40	499	✓
DFlash4-JUMP-MOBS	1.766	0.429×	7.4%	12.00	64.50	1.37	588	✓
DFlash5-FUSED-JUMP-MOBS	1.763	0.429×	7.7%	11.83	63.83	1.39	583	✓
DFlash6-Boltzmann	1.604	0.390×	5.8%	12.67	67.67	1.29	440	✓
DFlash6-BMOBS	1.853	0.450×	8.7%	11.50	61.50	1.43	413	✓
DFlash7-ACT	3.728	0.906×	52.4%	14.33	16.83	1.12	0	✓

Selected calibration settings

top_k: 8
jump_weight: 0.5
fused_weight: 1.0
boltzmann_temp: 0.1
bmobs_temp: 0.35
v7_margin_threshold: 3.0

Drafter diagnostics

Training position accuracies: 79.1%, 69.2%, 60.0%, 43.6%, 23.1%. Benchmark prompts are held out from the distillation seed list.

Method scope

DFlash2 uses exact top-k dynamic programming over a learned context-gated low-rank predecessor scorer. MOBS uses the same scorer with middle-out/gap selection. DFlash4 adds a separate +2/+4 jump head. DFlash5 reuses drafter hidden states for fused sparse anchors. DFlash6-Boltzmann uses deterministic adaptive-temperature Gumbel selection; BMOBS samples one middle anchor then fills with MOBS. DFlash7-ACT shortens low-confidence suffixes before Qwen verification. Every proposal is greedily verified and required to reproduce normal cached Qwen token-for-token.

Fidelity note: these are experimental mechanism-level adaptations in DFlash Mini Lab, not official upstream DFlash2–7 implementations or checkpoints.