v0.3.7 — Two models, both beating llama.cpp: Qwen3.6 +38%, Qwen3-30B +32%

sparkinfer now supports two models on a single RTX 5090 — and both outperform llama.cpp, with Qwen3.6 going from cold-start to 38% past llama.cpp in just 2 days. The moat is real and widening fast.

🏁 Two models. Both faster than llama.cpp.

Same RTX 5090, same Q4_K_M GGUF, warm & interleaved vs llama-bench:

Model	sparkinfer	llama.cpp
Qwen3-30B-A3B	~485 tok/s	~366 tok/s	+32%
Qwen3.6-35B-A3B	~387 tok/s	~275 tok/s	+38%

Qwen3.6 launched cold (llama.cpp itself has not been optimized for it yet), but sparkinfer closed the gap and pulled ahead — 38% past llama.cpp in 48 hours.

Performance — Qwen3.6 frontier 23 → 387 tok/s

• #223 Qwen3.6 Gated-DeltaNet hybrid decode on main — UD Q5_K quants, bf16 KV (@ai-hpc)
• #229 Gated-DeltaNet AR state update kernel with transposed state layout (@James-CUDA)
• #230 shared expert as coalesced GEMVs — 7.2× decode (23 → 167 tok/s) (@inference2026)
• #241 split-K bf16 dense projection GEMV for decode occupancy (@Dexterity104)
• #243 int8 dp4a MMVQ for Q5_K expert down-projection (@Dexterity104)
• #266 Q8 shared-expert MMVQ + MoE pipelining + F=512 gating + hd256 flash-attn — +15-20% decode (@jimcody1995)
• #269 Q8_0 on-read GEMV with split-K for native Q8_0 projection (@James-CUDA)

🔧 Eval & scoring fixes

• Difficulty anchor — Qwen3.6 scoring now uses Qwen3-30B llama.cpp ref as the maturity baseline (Qwen3.6 is not yet optimized in llama.cpp, so its own ref inflated difficulty_mult to 4.0×)
• Baseline measurement — dead code in same-box measurement unblocked, bash null-check vs value-check fixed, bot passes real measured values instead of hardcoded "0"
• Frontier ledger — safety clamp prevents Qwen3.6 deltas from leaking into the Qwen3-30B frontier
• Dual-model eval — Qwen3.6 scored as primary, Qwen3-30B as no-regression guard; both graded against live same-box origin/main
• Fair label anchoring — label tiers sized against llama.cpp reference, not the frontier

🔒 Scoring trust

• Copycat detection graduated — structural containment + Levenshtein + bigram cosine + DeepSeek LLM judge, with graduated warnings and auto-block on repeat offenders
• Same-box baseline — every PR graded against origin/main measured on the same box in the same session, eliminating hardware variance

Verified: RTX 5090, Qwen3.6 387 tok/s (128-tok), top-1 0.97 vs llama.cpp (KL ≈ 0.01).