--- license: other license_name: blackfrost-commercial base_model: - zai-org/GLM-5.3-BF16 base_model_relation: finetune pipeline_tag: text-generation library_name: transformers language: - en - zh tags: - glm - glm-5.3 - glm_moe_dsa - mixture-of-experts - moe - bf16 - de-risked - red-teaming - security-research - enterprise ---
![Blackfrost](https://cdn-uploads.huggingface.co/production/uploads/69a27f2d114e4ac9de4dafc7/xDTdhLFXmKlZOazFcvJ5S.jpeg) # GLM-5.3-DERISKED-BF16 ### Enterprise de-risked GLM-5.3 · 753B Mixture-of-Experts · BF16 master **Built by [Blackfrost](https://x.com/Blackfrost_AI) · Las Vegas, Nevada** ![BF16](https://img.shields.io/badge/Format-BF16-047857?style=for-the-badge) ![Manual access review](https://img.shields.io/badge/Access-Manual_review-1f2937?style=for-the-badge) ![Commercial](https://img.shields.io/badge/License-Commercial-b45309?style=for-the-badge) ![Price](https://img.shields.io/badge/Licensed_access-%24599-b91c1c?style=for-the-badge) ![Weights](https://img.shields.io/badge/Status-WEIGHTS_LIVE-047857?style=for-the-badge)
> **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. ## Licensed access — $599 USD This checkpoint is offered under a Blackfrost commercial license for **$599 USD**. Purchase, access, and enterprise licensing information is available from the [Blackfrost model catalog](https://redpillreader.com/models). 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`](https://huggingface.co/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`](https://huggingface.co/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 | ```text 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 `GlmMoeDsaForCausalLM` support. - 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 `1.0` and top-p `0.95`. ## Access and licensing The Blackfrost derivative is distributed under a separate commercial agreement. 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`](https://huggingface.co/zai-org/GLM-5.3-BF16/blob/main/LICENSE). **[Purchase or request licensed access — $599 USD](https://redpillreader.com/models)** For air-gapped deployment, enterprise support, evaluation services, or a custom build, contact [@Blackfrost_AI](https://x.com/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](https://x.com/Blackfrost_AI)