Skip to content
RCreddit.com·
Not on the current live radar

Qwen 3.8 Flash Next q4_k_m, 130k context, q8 cache on 16GB VRAM ann 64GB RAM, 15-20 t/s on 4080

AI summary

A user shared their experience running Qwen 3.8 Flash Next q4_k_m with 130k context and q8 cache on a system with 16GB VRAM and 64GB RAM, achieving 15-20 t/s on a 4080. Key factors for success include using the AtomicChat AD-4.27bpw Q4_K_M target, the Unsloth MTP head from a specific pr-mtp-fix branch, and the --spec-draft-cpu-moe flag. This configuration allows draft experts to reside in RAM, enabling hot experts to utilize the GPU, resulting in 16.5 tg / 350 pp at 131k with q8 KV.

Why this one

This report details specific, previously unshared configuration flags and branches that enable 130k context on 16GB VRAM, unlike general guides.

Time & source

Times shown in UTC

Display time zone: UTC

Local time zone unavailable; showing UTC.

IngestedOffset at this time: UTC+0Sep 24, 2026, 12:02 UTC

Ingested
Sep 24, 2026, 12:02
Source type
Dev community

Full text isn't available here.

Read at source →
Source·reddit.com