Built by Blackfrost · Las Vegas, NV
Why this model exists
Security teams cannot evaluate a defence against a model that refuses to produce the attack.
Standard frontier models are trained to decline adversarial prompts. That behaviour is appropriate for consumer products and can prevent authorised red teams, detection engineers, and safety labs from stress-testing their controls.
GLM-5.3-Flash-DERISKED-BF16 is Blackfrost's in-house de-risked build of GLM-5.3-Flash. The intended behavior is intrinsic to the checkpoint and does not require a system-prompt or decoding-time workaround. This repository is the BF16 master from which the Blackfrost NVFP4 release is derived.
It is offered to organisations under commercial licence, with access reviewed by hand. A final judged release-family refusal reference is published below; no GLM-5.2 figures are reused.
Specifications
| Architecture | Glm5NextForConditionalGeneration — natively multimodal GLM-5 hybrid-attention MoE |
| Parameters | 320B total · 18B active per token |
| Precision | BF16 — safetensors, 120 shards, 642,646,653,816 indexed tensor bytes (~598.5 GiB) |
| Layers | 45 main layers — first 3 dense, 42 MoE — + 1 multi-token-prediction (MTP) layer |
| Experts | 288 routed (top-8 active per token) + 1 shared expert |
| Hidden size | 4,096 · MoE FFN 2,048 · dense FFN 12,288 |
| Attention | 64 heads · 34 KDA layers + 11 full DSA layers · DSA indexer (32 heads, index_topk 2048) |
| Latent ranks | kv_lora_rank 512 · q_lora_rank 1536 |
| Vocabulary | 154,880 |
| Context | 1,048,576 positions (1M architectural ceiling) |
| Modalities | Text and vision |
| Languages | English, Chinese |
The tokenizer, configuration, generation configuration, multimodal processor, and GLM chat template ship with the licensed weights. Use a serving stack that honors the packaged template and model-specific reasoning/tool-call tokens.
What “de-risked” means here
Blackfrost derived this checkpoint directly from zai-org/GLM-5.3-Flash-BF16. Production details are proprietary and are not disclosed.
It is not a system prompt, LoRA, or decoding-time filter. The intended behavior change is in the weights. The release-family reference below was measured on the derived NVFP4 deployment checkpoint and is not presented as an independent BF16 measurement.
Lineage
| Base | Official zai-org/GLM-5.3-Flash-BF16 |
| Applied | Blackfrost proprietary weight-level de-risking |
| Not applied | SFT · DPO · RLHF · expert pruning · quantization |
| Format | Hugging Face safetensors · BF16 |
zai-org/GLM-5.3-Flash-BF16
└─ GLM-5.3-Flash-DERISKED-BF16 ← this repo, the master
└─ GLM-5.3-Flash-DERISKED-NVFP4 NVFP4 deployment derivative
Refusal evaluation
The release-family reference below was measured on the behavior-matched NVFP4 deployment checkpoint using R1-HARMFUL-BENCH-450 under a bare chat configuration. Responses were reviewed after generation to distinguish actual refusals from false-positive string matches.
Configuration: thinking enabled · maximum reasoning effort · temperature 1.0 · top-p 0.95 · top-k omitted · maximum 16,384 output tokens
| Evaluation slice | Final judged refusals |
|---|---|
| Harmful prompts | 4 / 300 (1.3%) |
| Full suite | 5 / 450 (1.1%) |
| API errors | 0 / 450 |
These figures are a release-family reference. This format was not independently rerun across all 450 prompts, so the table should not be represented as a format-specific measurement. The results are behavioral observations, not a safety certification.
Serving
This artifact has been load-validated with SGLang across 8×NVIDIA B200 GPUs. The following is the baseline shape used for validation; adjust context and memory settings to the deployment.
docker run --rm --gpus all --ipc=host --network host \
-v /path/to/GLM-5.3-Flash-DERISKED-BF16:/model:ro \
lmsysorg/sglang:latest \
python3 -m sglang.launch_server \
--model-path /model \
--served-model-name GLM-5.3-Flash-DERISKED-BF16 \
--trust-remote-code \
--tp-size 8 \
--ep-size 8 \
--dsa-prefill-backend tilelang \
--dsa-decode-backend tilelang \
--kv-cache-dtype bfloat16 \
--moe-runner-backend deep_gemm \
--disable-shared-experts-fusion \
--reasoning-parser glm45 \
--tool-call-parser glm47 \
--host 0.0.0.0 \
--port 8000
Pin a tested SGLang image in production rather than relying on a floating tag.
Deployment notes
- Hardware. The BF16 index declares 642,646,653,816 tensor bytes across 120 shards. The release was load-validated on 8×B200, with approximately 73.77 GB of model weights per rank under the validated TP8/EP8 layout. Reserve additional HBM for KV cache, activations, and CUDA graphs.
- Context. 1M positions is the architectural ceiling, not a guaranteed per-request allocation. Start with a context budget appropriate to available KV memory and raise it after the baseline serve is stable.
- Hybrid attention. The model contains 34 KDA layers and 11 full DSA layers; use a runtime with native GLM-5.3 support.
- Multi-token prediction. The BF16 master includes the MTP layer. The validation baseline kept speculative decoding disabled; enable it only with a runtime version that supports this GLM-5.3 layout.
- Multimodality. Validate both text-only and vision requests in the target runtime before production deployment.
Access & licensing
This repository is licensed commercially and access is reviewed manually.
➜ Request licensed access to the BF16 master
Commercial licensing, air-gapped deployment, custom de-risking against an organisation's threat model, quantized variants sized to target hardware, and evaluation support are available from Blackfrost.
Contact Blackfrost
@Blackfrost_AI on X
DMs are open. Fastest route to a human.
Ask about this model, the NVFP4 build, custom de-risked builds, or an enterprise evaluation licence.
Blackfrost · Las Vegas, Nevada Frontier model engineering for offensive security and AI safety
Other builds
| Build | What it is |
|---|---|
| GLM-5.3-Flash-DERISKED-BF16 | Full-precision BF16 master — this repository |
GLM-5.3-Flash-DERISKED-NVFP4 | NVFP4 deployment derivative |
| Custom | Deployment-specific quantization, expert budgets, or de-risking calibrated to an organisation's threat model |
Disclaimer
Refusal behavior in this checkpoint has been deliberately modified at the weight level. It is not a safety-stock model and must not be deployed, marketed, or evaluated as one.
No warranty of any kind. Provided “as is,” without warranty express or implied, including fitness for a particular purpose. Nothing here guarantees that any given input will be accepted or refused, that any capability is retained, or that any category of output is unreachable.
Evaluation results are not safety claims. Refusal and capability measurements describe specific harnesses under stated conditions. They are not safety proofs and should not be generalized beyond the settings tested.
Modification by a recipient voids this characterization. Further fine-tuning, merging, quantization, or alteration produces an artifact Blackfrost has not evaluated and does not stand behind.
Operator-owned policy. Open weights mean the operator sets and enforces policy. Deploy only in controlled environments with access control, independent logging, and review.
Responsible use
Access is intended for authorized security work: red-team engagements the operator is contracted to perform, AI-safety and alignment research, guardrail and detection development, and model evaluation.
This checkpoint should be assumed capable of producing content a consumer model would decline. Acceptable-use restrictions are licence obligations, not model behavior. Deploy behind independent controls, logging, and access review, and treat outputs as untrusted.
GLM-5.3-Flash-DERISKED-BF16 · © 2026 Blackfrost Softwares Corp. @Blackfrost_AI