Qwen 3.8 27B at a true 128K context with vision on a single 16GB RTX 5080 (NInfer)

On a single RTX 5080 16GB the author runs Qwen 3.8 27B through NInfer, a purpose-built engine, at a genuine 128K: both the configured context and the physically allocated KV capacity are 131,072, with Q4 group-64 KV, MTP-3 speculative decoding and CUDA graphs off at concurrency 1. The text model is mixed Q3/Q4/Q5 groupwise quantization, about 3.95 bits per weight. The strongest fully validated long-context result is a qualification on an exact 118,001-token prompt: 1,380.61 tok/s prefill, 71.57 tok/s decode, 44.74% MTP acceptance (2.31 tokens per round). A later mainline rerun on the same corpus matched it within run-to-run variation (1,378.85 / 71.44). The author is explicit that 71.57 tok/s belongs to that v1.2 118K run, not a fresh v1.3 measurement, and that a short-request 84.7 tok/s number is not comparable to it.

Vision coexists with the full 131,072 context and KV allocation through several memory fixes: the vision token/workspace budget is decoupled from the text context, GDN prefill convolution temporaries are released earlier, and vision weights are host-mapped instead of held in VRAM. At the 2048-token vision profile the card has only about 8.6 MiB free after startup, so it assumes a clean GPU; the author recommends 1792 for general use. A one-token test image (left red, right blue) was read correctly, and a deterministic six-second red/green/blue MP4 was returned in order (prefill 1,721.9 tok/s, decode 97.1 tok/s, MTP 100%). The engine adds rolling tool checkpoints for append-only agent loops and a default reasoning budget of 2048. The validated model artifact is published on Hugging Face as qwen3_8_27b.ninfer (16,461,267,456 bytes), with the runtime and artifact identities kept deliberately separate.