---
license: other
library_name: transformers
pipeline_tag: text-generation
base_model:
  - zai-org/GLM-5.3-BF16
base_model_relation: quantized
gated: manual
language:
  - en
  - zh
tags:
  - glm
  - glm-5.3
  - fp8
  - sglang
  - dflash2
  - de-risked
  - enterprise
---

![BLACKFROST RESEARCH](ASSETS/BLACKFROST-RESEARCH-BANNER.png)

# GLM-5.3-Derisked-FP8

Deployment-ready FP8 derivative of the full GLM-5.3 Mixture-of-Experts model.
The model retains the complete expert topology and native GLM tokenizer, chat
template, reasoning, tool-use, long-context, and MTP metadata.

This is a public, manually gated customer release. Weight access is reviewed by
Blackfrost Research and distributed under a separate commercial agreement.

## Specifications

| | |
|---|---|
| Architecture | `GlmMoeDsaForCausalLM` |
| Precision | Dynamic block-128 E4M3 FP8 |
| Layers | 78 main layers plus the native MTP layer |
| Experts | 256 routed experts, top-8 active, plus shared expert |
| Hidden size | 6,144 |
| Context ceiling | 1,048,576 positions |
| Source | `zai-org/GLM-5.3-BF16` |

Blackfrost's weight-level production process is proprietary and is not included
in this repository. No adapter or client-side system prompt is required for the
checkpoint's intended behavior.

## DFlash2 deployment

The included deployment kit enables SGLang DFLASH speculative decoding with the
separate revision-pinned `incoai/GLM-5.3-DFlash2` companion. Its weights are not
redistributed here. Review the companion's CC BY-NC-ND 4.0 terms before use.

See [`DEPLOYMENT/README.md`](DEPLOYMENT/README.md) and
[`DEPLOYMENT/LAUNCH.sh`](DEPLOYMENT/LAUNCH.sh).

## Use and limitations

Qualify capability, safety, long-context behavior, tool use, and speculative
acceptance for your workload before production. Operators remain responsible
for access control, monitoring, applicable-law compliance, and treating model
output as untrusted. The model is provided as is, without warranty.