REDDIT · REDDIT// 3h agoOPENSOURCE RELEASE

Qwen3.6 GGUFs add MTP, fixed templates

froggeric’s GGUF build packages Qwen3.6-27B with MTP heads, turbo KV-cache support, and a patched chat template for llama.cpp-compatible runtimes. The pitch is practical local coding: more speed, less memory pressure, and fewer template bugs when you run the model on consumer hardware.

// ANALYSIS

This is less a new base model than a deployment unlock, but that may matter more for local agentic coding than the original release itself.

–`--spec-type mtp` plus draft decoding is the main win here, with the author claiming roughly 2.5x faster generation on a Mac M2 Max
–`turbo4` KV cache is doing real work too, because long-context local models usually die on memory before they die on raw compute
–The fixed Jinja templates are not cosmetic; they address tool-call, developer-role, and thinking-tag incompatibilities that break in non-vLLM runtimes
–The release is still gated on a custom llama.cpp PR, so this is powerful but not yet a plug-and-play mainstream setup
–The hardware tables make the release feel actionable, especially for 24GB to 48GB Macs and GPUs where Qwen3.6-27B starts to look viable

// TAGS

llmquantizationlong-contextinferenceopen-sourceqwen3-6-27b-mtp-ggufqwen-fixed-chat-templates

DISCOVERED

3h ago

2026-05-06

PUBLISHED

5h ago

2026-05-06

RELEVANCE

9/ 10

AUTHOR

ex-arman68