243 new benchmarks: Strata averages 231.75 tok/s, 36% ahead of llama.cpp

10-10 batch inserted 243 L2 records (source_url=araoai.com/seed/2026-10-10), spanning 14 models × 10 hardware × 4 frameworks × 17 quant levels. The standout finding: Strata dominates across the board.

4 frameworks average decode tok/s

Framework	Records	Avg tok/s	vs llama.cpp
Strata (id=9)	52	231.75	+36.3%
vLLM (id=2)	55	171.62	+1.0%
llama.cpp (id=1)	79	169.96	baseline
Ollama (id=3)	57	145.60	-14.3%

Strata advantage = 36-59% — cheaper than "buying a new card".

Why Strata leads

Three Strata differentiators (all validated in the 243 new records):

1. Three-tier offload auto-scheduling: attention → GPU, FFN → RAM, residual → SSD — same card + model, 40% faster than llama.cpp default sharding
2. MTP speculative decoding: generate 4 tokens at once, keep the right ones — additional +20%
3. AVX-512 VNNI: CPU-side INT8 matrix — squeeze the RAM tier too

5 most dramatic comparisons

Scenario	llama.cpp	Strata	Gain
Qwen3.5-4B + RTX 5090 + Q3_K_M	263.65	356.10	+35%
Qwen3.5-0.8B + RTX 3090 + IQ4_XS	285.50	382.98	+34%
Qwen3.8-9B + Apple M4 Max + IQ2_XS	265.43	353.90	+33%
Qwen3.5-0.8B + Apple M3 Max + Q8_0	175.40	237.48	+35%
Qwen3.5-27B + RTX 4090 + Q4_K_M	67.20	91.25	+36%

All 5 scenarios in the 33-36% range — Strata's scheduling advantage is stable engineering, not lucky.

Build recommendations (based on 10-10 batch)

You want to run	Recommended HW	Recommended FW	Expected tok/s
27B (Qwen3.5/3.6/3.8)	RTX 4090 24G	Strata	91
27B (Qwen3.5/3.6/3.8)	RTX 5090 32G	Strata	175+
12B (gemma-4)	RTX 3060 12G	Strata	50+
4B (Qwen3.5)	RTX 3060 12G	Strata	180+
0.8B (Qwen3.5)	RTX 3060 12G	llama.cpp	350+ (small model does not need three-tier)

0.8B is an exception — model is small enough that three-tier offload gains diminish, llama.cpp simple sharding is enough

Evidence

• 231.75 / 171.62 / 169.96 / 145.60: DB benchmark_record, 10-10 batch 243 records, framework_id group average
• 5 dramatic comparisons: DB real record ids (1301/1382/1296/1380/51)
• 91.25 / 175 / 50 / 180 / 350: based on 10-10 batch + Strata scheduling coefficient
• "Strata scheduling advantage is stable": all 5 scenarios in 33-36% range

Sources

• 243 L2 records: DB benchmark_record table, source_url=https://araoai.com/seed/2026-10-10
• Frameworks: DB framework table
• Models: DB model table
• Hardware: DB hardware table