Malware analysis at machine speed. From suspicious sample to actionable intelligence in minutes. Discover Caronte

Beelzebub models / Release 001

Hugging Face Private · Beta testers only

Introducing

blz-1

Open weights. Uncensored. Built for blue teams and red teams.

Designed to remove learned refusal barriers. Run it on-premise with our open weights, or let Beelzebub manage it on our NVIDIA infrastructure, entirely hosted in Italy.

Private beta

Release 001

blz-1 is currently reserved for beta testers.

The weights are private on Hugging Face. Choose on-premise access or a fully managed deployment on Beelzebub’s own NVIDIA hardware, entirely hosted in Italy.

Security research.
On your terms.

Introducing Beelzebub models. A new space for open-weight AI, built around the people who defend systems and the people who test them.

It starts with blz-1: a foundation for security analysis, coding and experimentation. Choose self-hosted inference or a deployment fully managed by our team.

  • Open weights
  • Uncensored
  • Self-hosted inference
  • Fully managed option

Built for both sides
of security.

Different questions. Shared curiosity.
Explore where blz-1 can fit into your work.

Blue team

Understand the threat.

Explore the evidence. Question the assumptions. Bring a research model into your defensive workflow.

  • Threat analysis

    Explore suspicious behavior, connect observations and develop hypotheses for further investigation.

  • Detection engineering

    Work through detection ideas and draft logic to test against your own telemetry.

  • Security code review

    Examine code, reason about potential weaknesses and explore defensive improvements.

Red team

Challenge the assumptions.

Investigate how systems behave. Explore their boundaries. Build on a model you can run and adapt.

  • Vulnerability research

    Study code and system behavior, investigate weaknesses and form research hypotheses.

  • Authorized assessments

    Develop assessment ideas and reason through attack scenarios in environments you are authorized to test.

  • Experimentation

    Prototype security tools, explore agent workflows and evaluate your ideas in a controlled lab.

Know what you’re running.

Model size, hardware planning and token estimates for self-hosting blz-1. Fully managed deployments run on Beelzebub’s own NVIDIA hardware in Italy.

Model size
27B
Nominal parameters
Release formats
2 formats
Q6_K GGUF + BF16
Upstream context reference
262,144
Tokens · base-model limit, unverified for blz-1

Based on Qwen3.8-27B . Context figures below refer to the upstream architecture.

Choose your footprint.

Planning estimates

Q6_K GGUF

Quantized weights for local inference

≈ 22.1 GB

Estimated weights · 20.63 GiB

Free SSD space
30 GB
Available GPU / unified memory
32 GB
Host RAM budget
64 GB

BF16

16-bit model checkpoint

≈ 54.0 GB

Estimated weights · 50.29 GiB

Free SSD space
65 GB
Available GPU / unified memory
80 GB
Host RAM budget
128 GB

Starting budgets, not validated minimums. Hardware figures assume one sequence at an 8K–32K context. Keep room for the OS, runtime, cache and additional model copies. Unified-memory systems share a single pool; GPU and host budgets are not additive.

Weight estimates use nominal parameters × storage precision. Actual files can differ due to tensor precision, metadata and auxiliary components. See the Q6_K encoding , memory-sizing guidance and release files .

Context & memory

More context. More memory.

Input and generated output share the context window. The upstream reference is 262,144 tokens; the effective blz-1 limit still needs validation in your runtime.

Illustrative 16-bit KV cache · one sequence
Context tokens KV cache only
8K 8,192 tokens ≈ 0.5 GiB
32K 32,768 tokens ≈ 2 GiB
128K 131,072 tokens ≈ 8 GiB
256K 262,144 tokens ≈ 16 GiB

Assumes unchanged upstream attention dimensions. Excludes recurrent state, activations, workspace and other runtime allocations. Quantized weights do not imply a quantized KV cache.

Calculation & source

2 (K + V) × 16 attention layers × 4 KV heads × 256 dimensions × 2 bytes = 65,536 bytes per token. The table multiplies this by the context length.

Upstream model configuration

Generation speed

Token-speed estimate

Bandwidth-only estimate · not a blz-1 benchmark

An idealized calculation that reads the entire estimated weight footprint once per generated token, with all weights in device memory. It is not a prediction of actual throughput.

Illustrative memory bandwidth · output tokens/s
Memory bandwidth Q6_K GGUF tokens/s BF16 tokens/s
400 GB/s ≈ 18 ≈ 7
1,000 GB/s ≈ 45 ≈ 19
1,800 GB/s ≈ 81 ≈ 33

Assumes one sequence and one-token decoding, without batching, speculation or MTP. It excludes cache traffic, compute, dequantization and prompt processing. Actual throughput depends on the implementation; measure on your target hardware.

Calculation & measurement

Tokens/s in this simplified model ≈ memory bandwidth (GB/s) ÷ estimated weight size (GB). Actual bytes read can differ due to caching and tensor access patterns. Measure your hardware, runtime, context and concurrency to obtain practical generation speed.

Understanding inference benchmarks

Storage uses decimal GB; cache estimates use binary GiB. K denotes 1,024 tokens.

Your infrastructure.
Or ours. Your choice.

The same blz-1. Two ways to run it. Bring our open weights on-premise, or let Beelzebub manage everything on our NVIDIA infrastructure in Italy.

01 On-premise

Open weights

Our weights.
Your environment.

Run blz-1 entirely on your own hardware with our open model weights. Your team hosts and operates the full deployment, from runtime to tools.

  • Open weights for approved beta testers
  • Fully hosted on your own infrastructure
  • Your runtime, integrations and workflows
Request on-premise access

02 Fully managed

Hosted in Italy

Our infrastructure.
Your next breakthrough.

Focus on your research. Beelzebub handles deployment and day-to-day operation on NVIDIA GPU hardware we own and operate, entirely hosted in Italy.

  • All managed hardware hosted in Italy
  • Our own NVIDIA GPU infrastructure
  • Fully managed by the Beelzebub team
Discuss managed deployment

Available through our private beta. Tell us how you want to run blz-1.