---

license: apache-2.0

base_model:

- Qwen/Qwen3.6-35B-A3B

library_name: ds4

pipeline_tag: text-generation

tags:

- hebrus

- ds4

- gguf

- qwen3.6

- qwen35moe

- apple-silicon

- metal

- quantized

- expert-major

- imatrix

inference: false

---

<p align="center">

<img src="https://raw.githubusercontent.com/andreaborio/hebrus-site/main/public/og.png" alt="Hebrus Studio — Metal first. SSD when memory ends." width="900">

</p>

Qwen3.6-35B-A3B — Hebrus GGUF releases

This repository keeps the versioned Qwen3.6-35B-A3B releases for the

Hebrus inference engine. Hebrus began

as a fork of Salvatore Sanfilippo's

antirez/ds4. Historical DS4 names remain

only in filenames and serialized compatibility fields. Its routed MoE weights

are embedded in the GGUF as a self-describing ds4.expert_major.v2 store; no

sidecar, runtime repack, or second copy of the routed bank is required.

> Compatibility: existing filenames, the dsbox.json manifest, and the

> ds4.expert_major.v2 marker remain unchanged serialized contracts.

Versions

| File | Status | Runtime |

| --- | --- | --- |

| Qwen3.6-35B-A3B-DS4-ExpertMajor-v2-MLX-Affine4-G64.gguf | Current | Hebrus main at 73a332f or newer |

| Qwen3.6-35B-A3B-DS4-ExpertMajor-v2-Q4_K_S.gguf | Retired v2/Q4 evidence | Rejected by the current affine-only runtime |

| Qwen3.6-35B-A3B-DS4-ExpertMajor-v1-Q4_K_S.gguf | Legacy migration evidence | Retired Hebrus v1 runtime only |

| Qwen3.6-35B-A3B-ds4-Q4_K_S.gguf | Legacy canonical converter input | Generic GGUF tooling; rejected by current Hebrus inference |

Hebrus Studio and the root dsbox.json select only the affine4/group-64 ExpertMajor v2

artifact. The older files are

kept so existing links and reproducibility records remain valid; they are not

fallbacks and must not be auto-selected by current Hebrus.

This is a Hebrus-native GGUF extension, not a portable GGUF layout. Generic GGUF

loaders cannot reconstruct or execute its routed weights. Canonical Qwen GGUFs,

ExpertMajor v1 files, sidecar layouts, and non-Metal inference are not supported

by the qualified Hebrus runtime for this model. Use the original model or a

canonical quantization with another runtime when generic-loader compatibility

is required.

Artifact

| Property | Value |

| --- | --- |

| Base model | Qwen/Qwen3.6-35B-A3B |

| File | Qwen3.6-35B-A3B-DS4-ExpertMajor-v2-MLX-Affine4-G64.gguf |

| Size | 20,808,566,880 bytes (19.38 GiB) |

| SHA-256 | dd17266185833a9f05531ce366fd7284ddca1ed64aa3dcf06e321e8c72c9ea3d |

| Container | GGUF v3 with one embedded ds4.expert_major.v2 store |

| Routed geometry | 40 routed layers, 256 experts per layer |

| Routed payload | 18,119,393,280 bytes |

| Routed storage | MLX-compatible affine 4-bit groups: 32 packed bytes + BF16 scale + BF16 bias per 64 weights |

| Qualified hardware | Apple Silicon Metal, at least 16 GiB unified memory |

The affine conversion changes the physical representation consumed by Hebrus'

resident and SSD-streamed Metal kernels without introducing a sidecar or a new

container version. The same GGUF runs in AUTO resident or SSD mode. It is not a

portable llama.cpp, MLX, Ollama, or generic GGUF artifact.

Run with Hebrus

Use andreaborio/hebrus main at commit

73a332fef82a0bcdd567d17e0de17aa004cad85d

or a compatible descendant:

git clone https://github.com/andreaborio/hebrus.git
cd hebrus
make -j

./hebrus \
  -m /absolute/path/to/Qwen3.6-35B-A3B-DS4-ExpertMajor-v2-MLX-Affine4-G64.gguf \
  --ctx 8192

The embedded ExpertMajor v2 store, Metal backend, and residency policy are

selected automatically. No Qwen enable flag, sidecar path, pack path, or hash

environment variable is needed. AUTO uses SSD streaming when the resident

budget does not fit; the admission floor is 16 GiB unified memory.

Correctness and measured performance

Final qualification on an Apple M5 Pro 64 GiB measured:

| Prompt frontier | Resident prefill | Forced-SSD prefill |

| ---: | ---: | ---: |

| 128 | 523.72 t/s | 210.70 t/s |

| 2,048 | 1,661.18 t/s | 551.57 t/s |

| 8,192 | 1,421.90 t/s | 269.10 t/s |

| 32,768 | 877.34 t/s | 83.69 t/s |

The 128+16 lane measured 57.43 t/s resident decode and 23.94 t/s with a cold

321-expert SSD cache. Both modes produced the same 16 greedy token IDs. At 32K,

adaptive macro-prefill reduced expert reads from 60.776 GiB to 16.875 GiB

(-72.24%) with a -0.63% throughput delta. Actual graph allocation also passed

the model-declared 262,144-token maximum. These are bounded local measurements;

context, output length, thermals, storage, and memory pressure affect results.

Current limitations

• Apple Silicon Metal is the only qualified inference backend.
• At least 16 GiB unified memory is required; constrained hosts use SSD streaming.
• The file requires Hebrus commit 73a332fef82a0bcdd567d17e0de17aa004cad85d

or a compatible descendant implementing ds4.expert_major.v2 for

qwen35moe.

• Canonical Qwen GGUF inference, the retired ExpertMajor v2 Q4 payload,

ExpertMajor v1, sidecars, and non-Metal inference are deliberately rejected.

• A generic GGUF loader can inspect the outer container but cannot execute the

opaque routed store.

License and attribution

The base model is released by the Qwen team under the Apache 2.0 license. See

the upstream model card for model

architecture, training, safety, license, and attribution details.