Frontier weights now fit on a retail GPU, fully offline

The frontier of capable open-weight models is no longer a datacenter category. It is a desk category. A 125-billion-parameter model answering prompts at human reading speed, on a retail card, with no network call in the loop, is the same shift that put yesterday's server-class CPUs into laptops. The mechanism repeats: each consumer GPU generation gains enough VRAM to hold a more aggressively quantized model, and each quantization generation (here, IQ3_S, packing 125B into roughly 24GB) loses less quality than the last. The frontier arrives in pieces, and it never phones home.

The standard read is that capable AI is centralized, rented by the token, gated behind an API. The pattern underneath says otherwise. Strata's README documents the floor: an RTX 5070 with 12GB of VRAM and 64GB of system RAM, the kind of machine a working developer already owns, with 100 to 140 tokens per second claimed on an RTX 3090. A Hacker News commenter on different hardware, an RTX 4090 with 128GB of DDR5 and a Ryzen 7950x3d, just reported 124 tokens per second, the same order of magnitude from a second rig, which is what turns a maintainer benchmark into a reproducible result.

Two caveats are honest. The model name Qwen3.8-Flash-Next is not a recognizable mainline Qwen release, and IQ3_S's quality loss against unquantized weights is not measured. The pattern survives both: aggressive quantization plus consumer VRAM keeps moving the frontier down the hardware stack, pulling more capability off the API rails onto the local machine. The person at the desk gains; the assumption that capable AI has to be rented loses.