#2581 Community benchmark: RTX 3070 Ti 8 GB + Ryzen 5 7600X - Qwen3.6-35B-A3B abliterated Q5_K_S

Description

Results-only community benchmark submission (no engine changes). Full details and limitations are in the report README.

- **Hardware:** RTX 3070 Ti 8 GB (driver 617.14, 290 W power limit), Ryzen 5 7600X (AVX-512), 31.1 GiB RAM, XPG GAMMIX S70 BLADE 2 TB NVMe; Windows 11; CUDA 13.0 engine build.
- **Model:** `llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-GGUF` Q5_K_S (sha256 `137ea38c8d8761443dce77dd63ced173d6c18c1b3c88a6e836957c266191531a`), run in the UD-Q4_K_XL install slot; 32K context; MTP on.
- **Suite:** `tools/strata_bench.py` against a running server, thinking off, temperature 0 — chat 50.6 tok/s decode (46.6-60.7), 4K prompt 1189 tok/s, 16K 1100 tok/s, agent session 35.5 tok/s.

Notes and limitations (also in the README):

- Desktop apps (browsers, chat apps) were open; the box sits at its RAM edge (0.2-3.8 GiB free during the runs) — a second suite run under memory pressure showed chat medians ~39 tok/s with dips to ~14.
- This engine path (qwen36, 0.1.44.1) has no batch slots — `--clients 4` measured 36.5 tok/s total with clients waiting (p95 latency ~14 s) — and no conversation checkpoints (0 % reuse even for an identical repeated prompt).
- Extra measurements: 25K-token prompt 376 tok/s prompt / 14.7 tok/s decode / ~67 s TTFT; thinking-on decode ~49-54 tok/s; small correctness battery 5/5 (the code check was executed).
- All routed experts run from host RAM (none fit the 8 GB card); VRAM steady at 7,666 / 8,192 MiB.