GLM-5.3-SUPER SCARY ACORDING TO ANTHROPIC-BF16
Enterprise de-risked GLM-5.3 · 753B Mixture-of-Experts · BF16 master
Built by Blackfrost · Las Vegas, Nevada
Weights published: the verified BF16 checkpoint is available as 158 model shards plus three preserved MTP shards, with its configuration, tokenizer, generation settings, and native GLM chat template.
Intended audience
Access is intended for security firms, authorized red teams, AI-safety labs, guardrail and detection teams, and enterprise research groups operating in controlled environments.
Overview
GLM-5.3-DERISKED-BF16 is Blackfrost's full-precision, weight-level de-risked
build of the official zai-org/GLM-5.3-BF16
checkpoint. It is the master artifact for the Blackfrost GLM-5.3 release family.
The intended behavior is intrinsic to the checkpoint. It does not depend on a system prompt, adapter, or decoding-time filter. Production methods are proprietary and are not disclosed.
Specifications
| Architecture | GlmMoeDsaForCausalLM |
| Parameters | Approximately 753B stored parameters · Mixture-of-Experts |
| Precision | BF16 mixed with native FP32 metadata tensors |
| Artifact size | 1,506,659,919,872 indexed tensor bytes (~1.37 TiB) |
| Shards | 158 model shards + 3 MTP shards |
| Layers | 78 main layers + 1 multi-token-prediction layer |
| Experts | 256 routed experts · top-8 active per token · shared expert |
| Hidden size | 6,144 |
| Attention heads | 64 |
| Context ceiling | 1,048,576 positions |
| Vocabulary | 154,880 |
| Languages | English and Chinese |
Tokenizer, generation configuration, and the native GLM chat template are part of the release artifact. Deploy with a runtime that supports GLM-5.3's native reasoning and tool-call tokens.
Lineage
| Upstream | zai-org/GLM-5.3-BF16 |
| Blackfrost change | Proprietary weight-level de-risking |
| Not applied | Additional SFT · DPO · RLHF · expert pruning · quantization |
| Format | Hugging Face Safetensors · BF16 |
zai-org/GLM-5.3-BF16
└─ GLM-5.3-DERISKED-BF16 ← this repository
└─ GLM-5.3-DERISKED-NVFP4 deployment derivative
Validation status
The BF16 artifact passed structural and index-level integrity checks across all 59,585 tensors. The multi-token-prediction layer is present in the release.
No refusal-rate or capability score is claimed in this card before completion of the final judged evaluation. Results will be added only after qualification.
Deployment notes
- Plan capacity from the indexed 1.37 TiB weight footprint and reserve additional HBM for KV cache, activations, runtime workspaces, and CUDA graphs.
- Use a current serving stack with native
GlmMoeDsaForCausalLMsupport. - The architectural context ceiling is not a guaranteed per-request allocation; set the production context budget from available KV memory.
- Treat multi-token prediction as an optional optimization and qualify it independently in the selected runtime.
- GLM-5.3's recommended sampling baseline is temperature
1.0and top-p0.95.
License and support
Recipients remain responsible for compliance with applicable upstream terms,
export controls, local law, and their own authorization boundaries.
The upstream checkpoint remains subject to the
GLM-5.3 License.
For air-gapped deployment, enterprise support, evaluation services, or a custom build, contact @Blackfrost_AI.
Responsible use
This checkpoint is intended for authorized security testing, AI-safety and alignment research, model evaluation, and defensive engineering. It is not a safety-stock model. Operators must provide independent access control, logging, monitoring, and policy enforcement, and must treat model output as untrusted.
The model is provided as is, without warranty. Evaluation results describe specific test conditions and are not safety certifications or guarantees of behavior in another deployment.
GLM-5.3-DERISKED-BF16 · © 2026 Blackfrost Softwares Corp.
@Blackfrost_AI