32GB to 64GB of RAM took Qwen3.8-Flash-Next from IQ2_XS (2-bit) to Unsloth UD-IQ4_XS (4-bit) on a single 4090
I went from 32GB to 64GB of DDR5-5800 this week and moved my Strata setup off IQ2XS.
Proof1 community thread
Key points
- Box: RTX 4090 24GB, i9-14900K, 64GB DDR5-5800, Arch.
- The models sit on a SATA SSD (~550 MB/s), so anything Strata has to pull from disk while it generates hurts.
- In return, Strata's quant table has UD-IQ4XS matching Q80's top token 88.9% of the time, against 84.5% for IQ3XXS and 80.6% for Q20 (IQ2XS sits between those two and isn't in the table).
- The default route tail skip (STRATAROUTETAILSKIP=7) roughly doubles KL against Q80 on the Unsloth files in RAM-budget mode.
Sources (1)
- [1]32GB to 64GB of RAM took Qwen3.8-Flash-Next from IQ2_XS (2-bit) to Unsloth UD-IQ4_XS (4-bit) on a single 4090r/LocalLLaMA (top, daily) · Oct 17, 08:57 PM
I went from 32GB to 64GB of DDR5-5800 this week and moved my Strata setup off IQ2_XS.
Box: RTX 4090 24GB, i9-14900K, 64GB DDR5-5800, Arch.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 17, 2026Qwen 27B (IQ4_XS, Q8/Q4 KV) at full 262k context on a 16GB RX 9070 XT: 80+ t/s decode with Adaptive KV Streaming
- Oct 17, 2026Qwen Flash Next quant and temp comparison - performance vs task time vs quality
- Oct 17, 2026Qwen3.8-27B-Defiance-Uncensored-MTP-GGUF
- Oct 17, 2026Nemotron 3.5 Bolt, Muse Spark Open, and a Fast LFM2.5 24B
- Oct 16, 2026Improved Qwen3.8-Flash Chat Template (w/ Sharp Template terse option)
- Oct 16, 2026Qwen Flash Next potential SOTA Q2 nobody talks about