Qwen CYBER-FROST runs faster on PCs with little RAM, a new chat shows the model that answers it, and the web app's
home-screen icon opens the page.
- Qwen CYBER-FROST on RAM-short PCs: decode +26% (#654): on Linux, the experts that neither the VRAM nor the RAM
holds are now read straight from the model file, without the system's file cache, which such a PC cannot spare.
Maya checks a sample of experts against the file when it opens it and keeps the old path on a file system that does
not allow these reads (MAYA_EXPERT_DIRECT=0turns them off). On a PC where VRAM and RAM together cannot hold every
expert, the setup now suggests a context that leaves more of the VRAM to experts (a 32 GB card with 30 GB of RAM:
128K instead of 256K). Measured on a V100 32 GB with 30 GB of RAM; PCs that hold every expert are unchanged. - New chat (#662, thanks @needmorevram): a card shows the model that answers it (ready or busy, its context, the
last answer's speed, images, what it saved) with Switch model, and the ideas to start from stay below the message
box. On a phone the card is one row that opens the Models page. An answer's numbers (tokens, speed) are now behind
an info button next to Copy and Regenerate. - Web app on a phone's home screen (#656, thanks @needmorevram): the saved icon opens the page instead of a
not-found error; icons saved with an older version are sent to the page too.