b9850

ggml-org/llama.cppb9850Jun 30, 2026by github-actions[bot]same day as b9849

What changed since b9849

low risk

b9850 adds Qwen3Next model support and fixes DFLASH input handling for Qwen-Coder-Next.

New features

  • Qwen3Next architectures are now supported via registration of a new t_layer_inp tensor for attention normalization.

Improvements

  • Layer processing loop now assigns inputs more correctly, improving tensor wiring for newer Qwen variants.

Bug fixes

  • DFLASH no longer misbehaves for Qwen-Coder-Next because layer inputs are assigned properly in the processing loop.

Upgrade notes

No config or CLI changes are required; simply swap in the new binaries. Users running Qwen3Next or Qwen-Coder-Next (especially with DFLASH) should upgrade to pick up the corrected tensor wiring. The IPv6 URL parsing work from b9849 is unchanged in this build.

Analyzed Oct 31, 2026

Upgrading?

Release notes summary

Focuses on Qwen3Next model support by registering a new tensor for attention normalization and fixing input assignment issues in the layer processing loop to ensure correct DFLASH functionality.

Key Highlights

  • Qwen3Next model support
  • Added tensor for attention normalization
  • Fixed input assignment in layer processing
  • DFLASH fix for Qwen-Coder-Next

Release Notes (excerpt)

<details open> model : register t_layer_inp for qwen3next (#25141) * Fix input assignment in layer processing loop Fix DFLASH for qwen-coder-next * add line break Added tensor for attention normalization in Qwen3 model. </details> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - Ubuntu x64 (CPU) - Ubuntu arm64 (CPU) - Ubuntu s390x (CPU) - Ubuntu x64 (Vulkan) - Ubuntu arm64 (Vulkan)…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories