Built from Z.ai's original weights with our own importance matrix. The calibration corpora behind our builds are public.
## Highlights
- **753B parameters**, read from the published weight index. Mixture-of-Experts: 256 routed experts with 8 active per token plus 1 shared, 78 layers, the first 3 dense.
- **Same base as GLM-5.2.** Z.ai state every gain comes from post-training, and the two checkpoints carry an identical parameter count.
- **Coding**: per Z.ai a 50% improvement over GLM-5.2 on their in-house Z.ai Code Bench, and open-source SOTA on Terminal Bench 3.0 and Agents' Last Exam.
- **Emergent cyber capability**: Z.ai report state of the art on CyberGym for vulnerability discovery, with gains largest further up the exploitation chain, more than doubling GLM-5.2 on exploitation benchmarks.
- **Thinking budget** controlled by `reasoning_effort` with three levels, `low`, `high` and `max`. It defaults to `max`.
- **Sparse attention** with a learned indexer (`glm_moe_dsa`, top-2048 index), and one multi-token-prediction block in the checkpoint.
- **Bilingual**, English and Chinese.
- **Ships in FP8** upstream, so our GGUF base is converted from the FP8 checkpoint rather than from BF16.
- **Full imatrix quantization** with our public [calibration corpora](https://huggingface.co/datasets/AtomicChat/calib-corpora).
> [!NOTE]
> These GGUFs are **self-quantized from the original weights**, not a repack. The importance matrix keeps low-bit quants closer to the full-precision model.
> [!IMPORTANT]
> This is a 753B model. Even at four bits a full set of weights runs to several
> hundred gigabytes, and all of it has to fit in fast memory. In practice that
> means a large-RAM server or a serious multi-GPU rig, not a laptop and not most
> workstations.
> [!IMPORTANT]
> Always pass `--jinja` so the **GLM-5.3 chat template** is applied. Z.ai note that
> `clear_thinking` defaults to `false` in that template, so pass `clear_thinking=true`
> for chat scenarios.
## Model Overview
| Property | Value |
|---|---|
| Base model | `zai-org/GLM-5.3` |
| Parameters | 753B total, MoE with 8 routed experts plus 1 shared active per token |
| Layers | 78, the first 3 dense, 256 routed experts in the rest |
| Architecture | `GlmMoeDsaForCausalLM`, sparse attention with a learned indexer |
| Upstream precision | FP8 (`e4m3`, 128x128 block scales) |
| Vocabulary | 154,880 |
| Context length | 1,048,576 positions in the config; Z.ai evaluate up to 1M with context management |
| Languages | English, Chinese |
| Thinking | `reasoning_effort`: `low`, `high`, `max`. Defaults to `max` |
| This repo | GGUF quants (imatrix). The importance matrix we built is published here too |
Scores are Z.ai's published results for the base `zai-org/GLM-5.3`. Selected numbers from
their table: Terminal Bench 2.1 88.2, Terminal Bench 3.0 28.3, DeepSWE 1.1 66.9,
CyberGym 84.5, AutomationBench 48.2, HLE w/ tools 62.5, GDPval-AA v2 1769.
## Choosing a quant
| Quant | Size | Notes |
|---|---|---|
| `AD-IQ2_M` | — | Smallest usable. Aggressive low-bit for the tightest boxes. |
| `AD-IQ3_M` | — | Beats Q3 at similar size thanks to imatrix. Best low-memory pick. |
| **`AD-Q4_K_M`** | — | **Recommended default. Best balance of size, speed and quality.** |
| `AD-Q5_K_M` | — | A step up when the memory is there. |
| `AD-Q6_K` | — | Near lossless. |
| `AD-Q8_0` | — | Effectively lossless, reference quality. |
> [!TIP]
> Sizes fill in once the quants finish uploading. Pick the largest file that fits your
> memory with room for context.
`AD-` marks an Atomic Dynamic layout: bits are assigned per tensor role rather than left
to a preset, with the router and the shared expert held high and the routed experts
carrying the compression.
## Get started
> [!NOTE]
> GLM-5.3 uses the `glm_moe_dsa` architecture. The quants in this repo are still
> uploading, and running them needs a `llama.cpp` build that has landed GLM-5.3 support.
> Until then, [Atomic Chat](https://atomic.chat) is the easiest way to run it as support ships.
Run GLM-5.3 locally with:
- **[Atomic Chat](https://atomic.chat):** the easiest path. Open the app, search `AtomicChat/GLM-5.3-GGUF`, pick a quant, hit **Use this model**.
- **llama.cpp:** `llama-server -hf AtomicChat/GLM-5.3-GGUF:AD-Q4_K_M --jinja -c 8192`
- **Ollama:** `ollama run hf.co/AtomicChat/GLM-5.3-GGUF:AD-Q4_K_M`
- **LM Studio / Jan:** search the repo id, download any quant.
For the original FP8 weights rather than GGUF, Z.ai list SGLang, vLLM, TokenSpeed,
Transformers and KTransformers on the [base model card](https://huggingface.co/zai-org/GLM-5.3).
## Best practices
| Parameter | Value |
|---|---|
| temperature | 1.0 |
| top_p | 0.95 |
| reasoning_effort | `max` for benchmark reproduction, `low` or `high` to spend fewer tokens |
| clear_thinking | `true` for chat |
From Z.ai's evaluation settings (HLE w/ tools). Per-benchmark settings vary; see the base
model card for details.
## Run in llama.cpp
```bash
git clone https://github.com/ggml-org/llama.cpp
cmake llama.cpp -B llama.cpp/build -DBUILD_SHARED_LIBS=OFF -DGGML_CUDA=ON
cmake --build llama.cpp/build --config Release -j --target llama-cli llama-server
```
```bash
./llama.cpp/build/bin/llama-server \
-hf AtomicChat/GLM-5.3-GGUF:AD-Q4_K_M \
--jinja -ngl 99 -c 8192 -fa on
```
## How these were made
1. Download `zai-org/GLM-5.3` (original FP8 weights).
2. Convert to GGUF with a [llama.cpp](https://github.com/ggml-org/llama.cpp) build that supports the GLM-5.3 architecture (`glm_moe_dsa`, sparse attention with a learned indexer).
3. Build an importance matrix over our public [calibration corpora](https://huggingface.co/datasets/AtomicChat/calib-corpora).
4. Quantize the ladder with `--imatrix`, assigning bits per tensor role: router and shared expert high, routed experts carrying the compression.
## License
Released by Z.ai (zai-org) under the [GLM-5.3 license](https://huggingface.co/zai-org/GLM-5.3/blob/main/LICENSE). Quantized by Atomic Chat.