0.13.1 — Qwen3.8-Flash-Next, day-0

Qwen's newest flagship runs on your Mac the day it exists: qwen3.8-flash-next-4bit serves our own 4-bit MLX conversion of the 125B Qwen3.8-Flash-Next through a vendored qwen4_exp backbone. Also in this release: pull exactly one variant of a multi-variant repo, subfolder-quant repos download from the mirror again, and switching models mid-session got a real lifecycle.

• Qwen3.8-Flash-Next is served. rapid-mlx serve qwen3.8-flash-next-4bit runs our 4-bit conversion of Qwen's 125B MoE (~6B active parameters per token) through a vendored implementation of its qwen4_exp architecture, verified for parity against the reference implementation. Experimental, text lane, hermes tool calling; plan for a 128 GB+ Mac. The weights pull from Hugging Face — at 4-bit × 125B they are large enough that we deliberately serve them from the source rather than the mirror.
• Pull one variant, not the whole repo. rapid-mlx pull <repo> --bits 4 / --format mlx selects a single variant of a multi-variant repository instead of downloading every quantization it ships (#2338).
• Subfolder-quant repos download from the mirror again. Repositories that keep each quantization in its own subfolder (LFM2.5 among them) were hard-declining the CDN mirror and always falling back to Hugging Face; they now pull the exact subfolder from the mirror (#2279).
• Switching models mid-session got a real lifecycle. Replacing the resident model now quiesces admitted requests first, hands off before the old engine retires, and surfaces an honest abort reason to any stream it interrupts — a dictation session no longer dies because chat swapped assistants underneath it.
• The 8 GB Quickstart is safe again. The Mac app's Quickstart pick on the smallest Macs is back to a model that actually fits (#2432).