Blackfrost AI Open-Sources Official 'Cyber-Frost-Harness' Supporting 4 Release Formats

Blackfrost AI has added an official execution harness to the BF16, FP8, EXL3, and NVFP4 repositories of its CYBER-FROST 3.8 model and open-sourced it on GitHub

tau · October 6, 2026

#Blackfrost-AI #Cyber-Frost #LLM #EXL3 #NVFP4 #Quantization #OpenSource

Blackfrost AI Open-Sources Official 'Cyber-Frost-Harness' Supporting 4 Release Formats

On October 6, 2026, Blackfrost-AI officially launched 'Cyber-Frost-Harness', a dedicated runtime execution harness for CYBER-FROST-3.8, a specialized language model developed for authorized security research and vulnerability assessment. Along with integrating the harness into the model's official release repositories across four major formats—BF16, FP8, EXL3, and NVFP4—the team made the harness publicly available on GitHub as an open-source tool adaptable for any open-source language model.

Technical editorial graphic illustrating Blackfrost AI's Cyber-Frost-Harness architecture and its multi-quantization runtime support for BF16, FP8, EXL3, and NVFP4 formats

Image source: Blackfrost-AI / GitHub

Built on the Qwen4ExpForConditionalGeneration architecture, CYBER-FROST-3.8 is an artifact engineered for security professionals conducting authorized assessments, systems testing, and incident response. Providing a standardized execution harness addresses the fragmented runtime environments across varying precisions and formats, establishing a unified entry point for local security research and open-source model evaluation.

From BF16 to NVFP4: A Dedicated Runtime Harness for 4 Model Formats

The core technical strength of Cyber-Frost-Harness is its consistent support for four distribution checkpoints tailored to diverse hardware specifications and operational constraints.

Rather than being tied to a single precision, the harness provides dedicated configurations optimized for each compression method:

  • BF16 and FP8 Support: Native support for full-precision SafeTensors checkpoints in BF16 alongside optimized FP8 precision designed for modern inference workloads.
  • Official Collaborative EXL3 Checkpoint Support: Full execution support for CYBER-FROST-3.8-EXL3-SAGE-3.87bpw, developed in official collaboration with @ViC305 (vcruz305). This artifact is hosted on Hugging Face as a standalone checkpoint rather than an adapter.
  • NVFP4 Support: Integration with the 4-bit floating-point format optimized for modern NVIDIA GPU architectures, enabling efficient inference within constrained VRAM budgets.

This multi-format architecture allows researchers to choose the optimal checkpoint for their available GPU compute and memory footprint while using a single execution interface.

GitHub Open-Source Availability and Universal Model Extensibility

Blackfrost-AI chose not to restrict Cyber-Frost-Harness as a proprietary or model-exclusive utility, releasing the codebase on GitHub (Blackfrost-AI/Cyber-Frost-Harness).

As stated in the official announcement, users can "grab it on GitHub for any model," making the harness's runtime architecture easily adaptable across diverse open-source LLM environments.

  • Reusable Harness Architecture: Provides foundation code to load, evaluate, and benchmark different model architectures and quantized weight variants.
  • Local Research Pipeline Integration: Offers a reproducible pipeline for enterprise security teams and academic labs evaluating checkpoints in isolated on-premise environments.

By adhering to an open-source distribution model, the release benefits not only security specialists evaluating CYBER-FROST, but also the broader local LLM community exploring multi-quantization runtimes.

Spark TensorFold Compatibility and Platform Scope

On the infrastructure side, the development team confirmed compatibility with TensorFold for Spark environments.

Responding to community questions regarding execution frameworks and platform availability, Blackfrost-AI stated that the harness is already compatible with TensorFold on Spark:

  • TensorFold for Spark Compatibility: The maintainers confirmed in response to user inquiries that the harness already supports TensorFold for Spark environments.
  • Mac Platform Evaluation: When asked about broader platform support, the team stated they are exploring what can be done for Mac environments in the future.

At present, primary deployment remains centered on NVIDIA GPU and Spark TensorFold environments, while Mac support remains in the exploratory stage.

Active Research Phase and Practical Deployment Caveats

Security practitioners and machine learning engineers planning to integrate Cyber-Frost-Harness should keep several operational factors in mind:

  • Authorized Security Research Scope: CYBER-FROST-3.8 is specifically designed for authorized security practitioners conducting vulnerability discovery, security engineering, and incident response.
  • Active Quality Assessment Phase: The model distribution is currently in an active research evaluation stage, meaning teams should track ongoing updates and revisions from upstream maintainers.
  • Hardware Platform Guidance: While Mac support is being explored for future updates, current confirmed deployments are centered on NVIDIA GPU and Spark TensorFold setups.

For security engineers and open-source model researchers looking to unify multi-format quantization workflows under a single framework, the Cyber-Frost-Harness repository provides an accessible, modular foundation.

Sources