mere.run 0.45.0

This release adds native Qwen3.8 Flash Next inference with qualified mixed-Q2/Q4
and Q4 MLX distributions and verified MTP acceleration. It also makes LTX 2.5
generation substantially more memory-bounded and keeps model inventory fast on
large or externally linked stores.

Text

• added pinned mixed Q2/Q4 and Q4 MLX distributions of
  Qwen3.8-Flash-Next. The native Qwen4Exp text path implements four-stream
  hyper-connections, PLE n-gram embeddings, grouped normalization, sigmoid
  gated-delta output, and exact short-context attention. The mixed artifact is
  the qualified 128 GB Mac profile; the larger Q4 artifact is cataloged for
  machines with more memory. Both pulls require explicit acknowledgement of
  the Qwen Community License 1.0, and model pull now accepts the clearer
  --accept-license-terms spelling as an alias for
  --accept-model-license. Until QSA selection is integrated, Qwen4Exp fails
  closed when prompt plus requested generation exceeds its 2,048-token indexer
  budget and image input remains unqualified. The bundled one-layer MTP path is
  enabled by default for greedy decode, preserves exact target verification,
  and measured 1.43–1.78x faster decode on the mixed checkpoint; set
  MERERUN_Q35_MTP_SPECULATION=0 for a target-only comparison.

API

• fixed GET /v1/models latency on large or externally linked model stores by
  sharing one shallow fast inventory across chat and companion discovery instead
  of repeatedly validating every installed runtime.

Video

• reduced LTX-2.5 image-to-video memory pressure with a public, pinned,
  self-contained distilled distribution: the video pipeline remains BF16 while
  the bundled Gemma 4 language tower uses MLX affine Q4. One accepted model
  pull now installs every runtime component without downloading the original
  BF16 text tower or asking users to quantize locally. Runtime releases the text
  encoder and image VAE as soon as their conditioning tensors are materialized.
  Set MERERUN_LTX_MEMORY_TRACE=1 to report MLX allocator use at the major
  generation phases.
• fixed LTX convolutional-VAE decode admission to account for the large
  temporary Metal working set of late 3D convolutions. Standard distilled
  generation now automatically decodes long or high-resolution clips in
  overlapped temporal tiles instead of attempting a full-frame decode that can
  exhaust unified memory.

Included pull requests

• exact release range: #365,
  #366,
  #367, and
  #369.

Downloads

• macOS (Apple Silicon): MereRun-0.45.0.dmg
• Linux CUDA x86_64: tarball or Debian package

The Linux CUDA artifact was validated on an NVIDIA GeForce RTX 3080 Ti with Qwen3 ASR true inference, and the Debian package passed install, upgrade from 0.44.0, removal, and clean reinstall on Ubuntu 24.04.