128K hard limit on 64GB RAM, 32GB VRAM is too conservative

I am using IQ3_S quantization of the Qwen3.8-Flash-Next on Ubuntu. The setup reverts my 256K context length choice down to 128K. With full context used, the system shows >6GB of free RAM. 1.8GB more was needed for 256K instead of 128K. A manual change to the run script allowed me to run with full native context of 256K with zero issues. This may be different for windows systems or if other RAM-heavy programs are in use, but for my case there was no reason to force it to 128K.

Loving the about 2.5x speed in token gen and >6x on pp. 70%+ expert cache hit. Thanks!

Same here. I’m on a Windows system with 64 GB of RAM and 16 GB of VRAM, running the IQ3_S variant. The setup capped the context at 128K, but setting it to 256K manually in the .json seems to work fine, with just a slight performance hit.

You're right, and several people said the same. Setup should recommend a context, not force one. The next release changes that: the 128K pick stays the recommendation when setup estimates RAM is tight, but choosing 256K (or --context 262144) is kept, with a warning that it may run out of RAM under load. An earlier install's context is kept on update too.

Until then: set "--max-context", "262144" in the args of strata-<model>.json and start again. Keeping this open until the release.

tests: golden --yes / Enter configs on five PCs (#403 #406 #364 #384)

setup: recommend, never force - longer contexts, more GPUs, a bigger …

docs: the RAM rule recommends 128K, the low-RAM mode on several GPUs,…

Engine 0.1.33: a portable image encoder again (#411 #412), setup reco…

Released in 0.1.33: setup still recommends 128K when its RAM estimate is tight, but a longer context you pick (or --context 262144) is kept with a note, and an update keeps your earlier context. Thanks!