#2562 bench: community report, 3x GPU (RTX 5060 Ti + 2080 Ti + 3080 Ti), Qwen 125B MoE IQ3_XXS

Description

### Summary

Community benchmark report measuring **Qwen3.8-Flash-Next** (125B MoE, IQ3_XXS, 3.4 bpw) running across a heterogeneous 3-GPU desktop setup using Strata v0.1.44 with MTP speculative decoding (`--spec 4`, `--spec-min-p 0.65`).

### Hardware and Software
- **GPUs:** 3 NVIDIA GPUs (39.8 GiB total VRAM)
  - GPU 1: NVIDIA GeForce RTX 5060 Ti (16 GiB, `sm_120`, Blackwell)
  - GPU 0: NVIDIA GeForce RTX 2080 Ti (11 GiB, `sm_75`, Turing)
  - GPU 2: NVIDIA GeForce RTX 3080 Ti (12 GiB, `sm_86`, Ampere)
- **Host CPU:** Dual Intel Xeon E5-2683 v4 (32 cores / 64 threads)
- **RAM:** 185 GB DDR4
- **OS / Drivers:** Linux 7.2.6-generic, NVIDIA Driver 615.71.09, CUDA 13.4
- **Engine:** Strata v0.1.44 (commit ed4773eb), native unified kernel build (`sm_75`, `sm_86`, `sm_120`)

### Configuration
- **Model:** `Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_XXS`
- **Context:** 262,144 tokens (`--kv k8v4`)
- **Parallel Pipeline:** `--layer-split 17,29` across the 3 GPUs
- **Speculative Decoding:** `--spec 4 --spec-min-p 0.65 --mtp rt`
- **Vision:** `--vision` sidecar enabled

### Results Summary
- **Short Chat Decode:** **104.0 tok/s** median (89.9 – 117.7 tok/s), TTFT **0.45 s** (450 ms)
- **4K Prompt Prefill:** **1,033.8 tok/s** median, TTFT **3.91 s**
- **16K Prompt Prefill:** **2,046.8 tok/s** median, TTFT **7.88 s**
- **32K Prompt Prefill:** **2,788.0 tok/s** median, TTFT **11.55 s**
- **Agent Session:** **108.9 tok/s** median decode (up to **154.8 tok/s** peak), **68.7%** cache hit rate, **4** tokens re-read per session across 6 turns.

All files are formatted per `docs/COMMUNITY_BENCHMARKS.md`, with personal paths, hostnames, and IP addresses automatically redacted.