andreaborio/Qwen3.6-35B-A3B-Hebrus-GGUF

Qwen3.6-35B-A3B — Hebrus ExpertMajor v2

This repository contains the Stable and Beta Qwen3.6-35B-A3B profiles for andreaborio/hebrus, a source-built Apple Metal inference engine with bounded SSD expert streaming.

Both artifacts embed their routed MoE weights in a checksummed ds4.expert_major.v2 store. Hebrus uses one Qwen graph, tokenizer, session, attention, Gated DeltaNet, KV, routing, resident/SSD policy, and prefill scheduler for both accepted profiles. Only the admitted physical weight codec differs.

These are Hebrus-native GGUF extensions. They are not portable to generic llama.cpp, MLX, Ollama, or hosted Hugging Face inference.

Published profiles

Both canonical files are pinned at immutable Hugging Face revision e002665becd2db618897effb213030fdf92e7e98.

Profile	File	Status	Bytes	SHA-256	Minimum unified memory	Qualified boundary
MLX Affine4 G64	Qwen3.6-35B-A3B-Hebrus-ExpertMajor-v2-MLX-Affine4-G64.gguf	Published · Stable · Recommended; selected by qwen-v2	20,808,566,880	dd17266185833a9f05531ce366fd7284ddca1ed64aa3dcf06e321e8c72c9ea3d	16 GiB	On 16 GiB, guarded SSD through a 131,072-token prompt plus 128 decode tokens; larger windows require separately qualified higher-memory profiles
Q2_K_XL	Qwen3.6-35B-A3B-Hebrus-ExpertMajor-v2-Q2_K_XL.gguf	Published Beta · Opt in · Not recommended; selected only by qwen-q2-beta	12,290,632,032	30c22f70aff0f05986b517ee4ad8fef554a1b5aab6971c9ca09f999566d30143	64 GiB	Resident and SSD evidence through exactly 32,768 tokens; no full-window qualification

The Q2_K_XL embedded routed-payload SHA-256 is ccc3fbc2405d1dd73f8ac15741b0277514de4f46b80818531297ea9ffa0c6a3c. It is 40.93% smaller than the Stable file, but its smaller size does not lower its published 64 GiB minimum or make it the recommended profile.

The format-compatibility floors are 73a332fef82a0bcdd567d17e0de17aa004cad85d for Affine4 and 42e2fec2a7dbb14a42e7a5612dfec00e33d443ca for Q2_K_XL. Those commits describe artifact decoding only; they are not substitutes for a current runtime containing all accepted safety fixes. The commands below pin the newer source snapshot used to verify this card.

Quick start: Stable and recommended

Requirements:

• an Apple Silicon Mac with at least 16 GiB unified memory;
• Xcode Command Line Tools;
• enough local SSD space for the 20,808,566,880-byte artifact;
• the official Hugging Face CLI.

git clone https://github.com/andreaborio/hebrus.git
cd hebrus
git checkout v0.3.0

xcode-select --install
python3 -m pip install -U huggingface_hub hf_xet
make -j

./download_model.sh qwen-v2

./hebrus \
  -m gguf/Qwen3.6-35B-A3B-Hebrus-ExpertMajor-v2-MLX-Affine4-G64.gguf \
  --ctx 8192 \
  -p "Explain bounded SSD expert streaming in one paragraph."

The downloader pins the immutable Hub revision and verifies the exact byte count and SHA-256. The repository is public and ungated, so downloading does not require a Hugging Face token.

Use normal AUTO startup. On the 16 GiB tier AUTO resolves to guarded SSD. Do not force resident mode or a manual expert-cache size.

Opt-in Beta

The Q2_K_XL profile has a separate selector and an intentionally narrower contract:

./download_model.sh qwen-q2-beta

./hebrus \
  -m gguf/Qwen3.6-35B-A3B-Hebrus-ExpertMajor-v2-Q2_K_XL.gguf \
  --ctx 32768 \
  -p "Explain bounded SSD expert streaming in one paragraph."

Use it only on an Apple Silicon Mac with at least 64 GiB unified memory. It is qualified only through 32K and makes no 262K/full-window claim.

Historical and rejected files

The repository retains pre-Hebrus DS4 basenames as historical, byte-identical objects. New commands and downloads use the canonical Hebrus filenames above.

The older Qwen3.6-35B-A3B-DS4-ExpertMajor-v2-Q4_K_S.gguf is negative-only, not a fallback. It is 20,808,566,880 bytes and has SHA-256 d7c43a6388ec20e6fe5530850350f96fdb0ac37c5ce36d3e5f92b172c447f56b. ExpertMajor v1 and the canonical Qwen3.6-35B-A3B-ds4-Q4_K_S.gguf converter input are also rejected by current inference.

Compatibility and limits

• Apple Silicon Metal is the only qualified production backend.
• CPU is reference/debug only. CUDA and ROCm sources are absent, and distributed inference is retired.
• Hebrus admits complete, exact tensor/tokenizer/ExpertMajor inventories and fails closed for malformed or crossed profiles.
• The 16 GiB qualification belongs only to the Stable Affine4 artifact. It does not imply lower-memory support for Q2_K_XL, DeepSeek, or GLM.
• Qwen currently supports the Hebrus server's /v1/chat/completions path, not every protocol implemented for other model families.
• Performance depends on the exact artifact, runtime revision, context, prompt, mode, memory pressure, storage state, and thermals.
• inference: false is deliberate: the Hub hosted inference widget does not support these runtime-specific artifacts.

The machine-readable authority is the Qwen release contract, and the current hardware/mode authority is the runtime support contract.

License and attribution

The upstream Qwen/Qwen3.6-35B-A3B repository declares the Apache 2.0 license. The Qwen team retains the base-model copyright; the embedded ExpertMajor v2 layout and current runtime are maintained by Hebrus.