| Server | http://strata/v1/chat/completions |
| Serving GPU (verified 2026-10-13) | GET /metrics -> hardware_static.gpu_name 'NVIDIA GeForce RTX 3080 Ti', gpu_count 1, 12,288 MiB total / 12,194 MiB used, 261 W of a 370 W limit, PCIe gen3 x8. Confirmed on the host: card1 = vendor 0x10de device 0x2208 driver nvidia (RTX 3080 Ti); card0 = vendor 0x8086 device 0xe223 driver xe (Intel Arc Battlemage — installed, but NOT the serving GPU). nvidia-smi on strata lists exactly one GPU: index 0, RTX 3080 Ti, 12288 MiB. |
| Model | qwen3.8-flash-next-iq3_xxs |
| Composite score | 9.8 (9/9 tests) |
| Speed | ~56 tok/s (simple), ~60 tok/s (games) |
| Architecture | 125B total parameters, 6B activated per token, 51B n-gram embedding |
| GPU residency (verified post-run via /metrics) | Model fully loaded: 462 MiB VRAM free, expert cache 2,547 MiB resident, KV int8 @ 131K ctx, ~20% expert traffic over PCIe (RAM-backed expert slots) |
| Type | Ultra-sparse Mixture-of-Experts (MoE) with vision encoder |
| MoE details | 512 experts, 10 routed + 1 shared activated per token |
| Context | 262,144 native, extensible to 1,000,000 tokens |
| Run window | 2026-10-12 19:24 → 21:47 |
· [assistant] 60 tok/s on game tests is comparable to qwen3.5-9b's 56 tok/s on the GTX 1080 — notable given this is a 125B-total/6B-activated ultra-sparse MoE running on the box's RTX 3080 Ti (12 GB), with the expert cache spilling to system RAM over PCIe. All outputs verified self-contained. (inference from game tests)
· Hardware label corrected 2026-10-13: this row previously read 'Intel Arc B70 SYLLAMA'. Strata's own /metrics names the serving GPU as the NVIDIA RTX 3080 Ti, and the Arc (vendor 0x8086, driver xe) is installed but unused for serving. The older strata-era rows (qwen3.8-27b, swift-1.5-iq2-s, frognano-4b-*) still carry 'Arc B70' claims inherited from an env record written that week — those describe a past run and have NOT been re-verified against current metrics.