PolyStrata on a
12 GB gaming GPU
RX 6700 XT · Ryzen 5 3600 · 32 GB system RAM · native HIP gfx1031
A major throughput improvement, with different quality tradeoffs by model. Six model families, seven quality profiles, and a separate same-weight speed control.
Same weights. Faster measured configuration.
Qwen3.6 UD-Q4_K_M, 65,536-token configured window. The 22.66 GB model files have identical SHA256 hashes.
| Metric | Historical Strata | PolyStrata control | Ratio |
|---|---|---|---|
| Short-chat median | 27.5 tok/s | 79.7 tok/s | 2.90× |
| 16K prompt median | 678.0 tok/s | 1,021.3 tok/s | 1.51× |
| 32K prompt median | 543.3 tok/s | 816.6 tok/s | 1.50× |
Runtime, draft implementation and settings differ. Tail-skip is disabled for the new control. This is not an isolated kernel A/B test, and the speed-only control has no custom quality score.
Model results: speed and quality stay separate
64K configured windows. Practical: 24 tasks × 3, fixed weighted rubric out of 100. HumanEval+: all 164 problems, plus-test passes. Empty finals and output-limit finishes count across 236 practical/coding requests per profile.
| Model / exact quant | Chat tok/s | Practical /100 | HumanEval+ | Empty finals | Output-limit finishes |
|---|---|---|---|---|---|
| Tiel Coder 35B-A3BUD-Q3_K_XL | 68.7 | 98.41 | 136/164 | 0 | 1 |
| Qwen3.6 35B-A3BUD-Q3_K_M | 79.9 | 96.61 | 120/164 | 0 | 0 |
| Ornith 1.5APEX-I-COMPACT | 69.2 | 96.12 | 141/164 | 0 | 0 |
| Gemma4 26B-A4B · thinking disabledUD-Q4_K_M | 72.8 | 95.28 | 153/164 | 0 | 1 |
| MiniCPM-V 4.7mix_q3gud_im | 59.2 | 91.33 | 155/164 | 5 | 1 |
| LFM2.5 24B-A2BUD-Q4_K_XL | 80.5 | 91.20 | 148/164 | 0 | 2 |
| Gemma4 26B-A4B · originalUD-Q4_K_M | 73.9 | 49.73 | 100/164 | 90 | 74 |
The previous Qwen Q4_K_XL scored 98.34 practical and 130/164 coding; Coder IQ1_M scored 95.35 and 152/164. Small differences on this fixed suite do not establish a general winner. MiniCPM was text-only; vision was not benchmarked.
Allocation is not retrieval
All six models passed 8K/32K/64K/80K/128K admission. The table below shows actual input tokens in three separate queries at depths 10/50/90. These are retrieval tests, not full quality-suite runs at 128K.
| Model | Allocated | Actual input tokens | Recall | Method |
|---|---|---|---|---|
| Gemma4 26B-A4B | 81,920 | 74,247–74,248 | 3/3 | token-calibrated follow-up |
| Gemma4 26B-A4B | 131,072 | 126,361–126,362 | 3/3 | token-calibrated follow-up |
| LFM2.5 24B-A2B | 81,920 | 76,878–76,879 | 0/3 | original upstream needle |
| LFM2.5 24B-A2B | 131,072 | 127,249–127,251 | 3/3 | original upstream needle |
| MiniCPM-V 4.7 | 81,920 | 74,200–74,201 | 3/3 | token-calibrated follow-up |
| MiniCPM-V 4.7 | 131,072 | 126,449–126,450 | 3/3 | token-calibrated follow-up |
| Ornith 1.5 | 81,920 | 74,200–74,201 | 3/3 | token-calibrated follow-up |
| Ornith 1.5 | 131,072 | 126,449–126,450 | 3/3 | token-calibrated follow-up |
| Qwen3.6 35B-A3B | 81,920 | 74,200–74,201 | 3/3 | token-calibrated follow-up |
| Qwen3.6 35B-A3B | 131,072 | 126,449–126,450 | 3/3 | token-calibrated follow-up |
| Tiel Coder 35B-A3B | 81,920 | 74,378–74,379 | 3/3 | token-calibrated follow-up |
| Tiel Coder 35B-A3B | 131,072 | 126,627–126,628 | 3/3 | token-calibrated follow-up |
LFM caveat:0/3 at 8K,0/3 at 16K,3/3 at 32K,0/3 at about 76,879 actual tokens, then3/3 at about 127,250. Misses produced explanatory text under the fixed 40-token budget instead of the requested word. Original oversized HTTP400 attempts from the other models are preserved; corrected queries are separately labeled.
Official upstream speed suite
Unmodified tools/strata_bench.py, three runs per test; short-chat medians use nine requests. A dash means not measured at that tier. These rows are speed measurements, not custom quality tests.
| Profile | Allocated tokens | Chat tok/s | 16K prompt tok/s | 32K prompt tok/s |
|---|---|---|---|---|
| gemma4-ud-q4_k_m-reserve1024 | 32768 | 73.5 | 988.3 | — |
| gemma4-ud-q4_k_m-reserve1024 | 65536 | 73.9 | 985.3 | 819.1 |
| lfm25-ud-q4_k_xl-reserve1024 | 32768 | 84.0 | 1501.2 | — |
| lfm25-ud-q4_k_xl-reserve1024 | 65536 | 80.5 | 1490.1 | 1120.6 |
| minicpmv-mix_q3gud_im-reserve1024 | 32768 | 59.2 | 1008.2 | — |
| minicpmv-mix_q3gud_im-reserve1024 | 65536 | 59.2 | 1006.1 | 790.5 |
| ornith-apex-i-compact-reserve1024 | 32768 | 68.5 | 989.2 | — |
| ornith-apex-i-compact-reserve1024 | 65536 | 69.2 | 991.5 | 781.9 |
| qwen36-ud-q3_k_m-reserve1024 | 32768 | 75.8 | 1104.0 | — |
| qwen36-ud-q3_k_m-reserve1024 | 65536 | 79.9 | 1103.4 | 850.9 |
| qwen36-ud-q4_k_m-control-reserve1024 | 32768 | 80.7 | 1019.4 | — |
| qwen36-ud-q4_k_m-control-reserve1024 | 65536 | 79.7 | 1021.3 | 816.6 |
| tiel-ud-q3_k_xl-reserve1024 | 32768 | 71.0 | 1077.7 | — |
| tiel-ud-q3_k_xl-reserve1024 | 65536 | 68.7 | 1088.3 | 838.7 |
Stability, resources and sandbox
The initial default-reserve trial hit GPU memory-access faults. Explicit 1 GiB reserve profiles completed with no recorded GPU memory-access faults and no guest OOM or OOM-kill events. At 64K, sampled minimum free VRAM was about 992–1,015 MiB. Five-second sampling can miss brief peaks; this is not a guarantee of indefinite production stability.
The maximum sampled GPU sensor was 93°C, including hotspot sensors. Guest cgroup peaks include filesystem cache and are not anonymous model-RAM requirements. Inference ran nonroot/offline/read-only with capabilities dropped and seccomp filtering; generated code ran in a separate offline scorer after inference stopped.
Data and methodology
Structured benchmark data · Release verification · Full results CSV
Local hardware was used. No cloud inference/scoring or quant creation was billed in this experiment. Electricity, depreciation, labor and total project cost were not measured; do not interpret this as zero total cost.