---

base_model: zai-org/GLM-5.2

library_name: ds4

license: mit

language:

• en
• zh

pipeline_tag: text-generation

tags:

• gguf
• hebrus
• ds4
• glm-dsa
• glm-5.2
• apple-silicon
• metal
• q2-k
• moe
• expert-major

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>

GLM-5.2 — Hebrus GGUF releases

This is the definitive versioned GLM-5.2 model container 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. The current release is

one qualified ExpertMajor v2 GGUF. The older chunked diagnostic artifact and

its sidecar are retained only as historical experimental evidence.

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

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

Versions

| Artifact | Status | Runtime |

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

| GLM-5.2-DS4-ExpertMajor-v2-Q2_K.gguf | Current | Hebrus main at 57acfd4 or newer |

| glm52-ds4-native-64g.gguf.part-000 through part-005 | Legacy experimental diagnostic | Retired preview branch only; rejected by current Hebrus inference |

| sidecars/selected-hotlist-n32-pack2600/ | Legacy sidecar experiment | Retired; never selected by current Hebrus |

Hebrus Studio and the root dsbox.json select only the single-file ExpertMajor v2

release. The legacy files remain available for reproducibility and historical

links; they are not fallback formats. Current Hebrus provides no compatibility

path for the old chunked canonical layout or its sidecar.

The current file embeds the routed MoE bank once as

ds4.expert_major.v2. It needs no sidecar, startup repack, or second copy of

the routed weights. It is intentionally Hebrus-only: llama.cpp, MLX, Ollama, and

other generic GGUF loaders cannot execute it.

Current artifact

| Property | Value |

| --- | --- |

| Base model | zai-org/GLM-5.2 |

| File | GLM-5.2-DS4-ExpertMajor-v2-Q2_K.gguf |

| Size | 262,147,193,504 bytes (244.14 GiB) |

| SHA-256 | 7f5017e3076e706c78f2a5322b035a9e2f6519c65ff5b6be8b2d91aeff61505d |

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

| Tensor inventory | 1,297 physical tensors; 1,524 logical tensors reconstructed by Hebrus |

| Routed geometry | 76 routed layers, 256 experts per layer, top-8 |

| Expert record | gate + up + down, 12,386,304 contiguous bytes |

| Routed payload | 240,987,930,624 bytes (224.44 GiB) |

| Routed payload SHA-256 | 122dcd3185037a3c0bcc8abab18ef61203ecfbfd2b25d02bd01e49c48778dfa9 |

| Qualified hardware | Apple Silicon Metal, at least 64 GiB unified memory |

The v2 conversion changes routed-weight order and the GGUF tensor directory

without requantizing values. Non-routed payloads are copied from the

byte-identified source and routed weights are stored once.

Run with Hebrus

Use andreaborio/hebrus main at commit

57acfd408a3154851a0c59be432904300abb3b6c

or a compatible descendant:

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

./hebrus \
  -m /absolute/path/to/GLM-5.2-DS4-ExpertMajor-v2-Q2_K.gguf \
  --ctx 8192

The runtime automatically selects Metal, SSD residency, the GLM profile,

contiguous-record expert reads, compact DSA KV, and the qualified cache policy.

No Metal, SSD, cache, preload, full-layer, sidecar, or ExpertMajor flag is

needed.

Measured performance

The release lane used an Apple M5 Pro with 64 GiB unified memory, local APFS

storage, context 8192, an exact 288-token prompt, and 32 greedy output tokens:

| Runtime arm | Prefill | Decode |

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

| Clean agent-friendly v2-only refactor | 11.89 t/s | 1.84 t/s |

| Post-unification v2-only gate | 11.20 t/s | 1.82 t/s |

| Earlier public-command gates | 10.63-10.91 t/s | 1.77-1.79 t/s |

| Prior rested-storage qualification, median | 11.075 t/s | 1.900 t/s |

Every accepted arm produced the same deterministic continuation. These are

bounded local measurements, not universal throughput guarantees.

The final compact-indexer correction was also exercised at 32K on the same 64

GiB host. The prose lane measured 44.73 t/s prefill, 1.87 t/s overall decode,

and 472.624/621.302 ms TPOT p50/p95; p50 corresponds to about 2.12 t/s steady

decode. A separate security/coding prompt measured 45.53/1.33 t/s because its

expert-cache hit rate fell from 36.90% to 13.15%. Both runs completed with zero

new swapout and exact retained evidence. The original baseline fails the 32K

transition before logits, so this is a correctness qualification rather than a

speedup claim.

Legacy experimental artifact

The retained diagnostic source reconstructs to

glm52-ds4-native-64g.gguf, 262,147,189,920 bytes, SHA-256

99e18cf31620845f6ff8e08ec8074865b613a3684383fa53ebb93e77260d7dd5.

It is stored as six part-000 through part-005 files with checksum manifests

and reconstruct.sh. The optional 32,204,390,400-byte selected-hotlist sidecar

is also historical. Neither layout is supported by current Hebrus inference.

These files document the early CPU/preview and selected-hotlist bring-up that

preceded the embedded ExpertMajor v2 store. Keep them for reproducibility; do

not use their old startup commands with current main.

Limitations

• Minimum qualified memory is 64 GiB unified memory.
• One complete local ExpertMajor v2 file and Apple Metal SSD streaming are

mandatory.

• 8K is the normal 64 GiB starting point and 32K is qualified but slow to test;

no 65K/100K GLM claim is made because this release does not change attention,

KV geometry, RoPE, or context scaling.

• MTP/speculative decoding remains disabled because measured acceptance did

not repay its additional expert I/O.

• Performance depends on storage state, memory pressure, context length,

routing locality, and thermals.

License and attribution

GLM-5.2 is released by Z.ai under the MIT license. See the upstream model card

for architecture, training, safety, and attribution details. The historical

runtime work credits antirez/ds4 and

andreaborio/hebrus.