BOLTBREAKER

A less refusal-prone Nemotron 3.5 Bolt, with its vision weights left intact.

BOLTBREAKER is an independent OBLITERATUS-derived version of NVIDIA Nemotron 3.5 Bolt 74B-A9B. It uses NF4 quantization for experts and retains BF16 residual output matrices and vision components. Built and tested on a DGX Spark / GB10.

The reason to try it: in our 169-prompt comparison, it refused 1 of 45 benign prompts, versus 3 of 45 for the stock Super-VL-120B baseline. Cyber/pentest refusals were 3/30 versus 16/30. These are observations from a small test against a different stock model, not a claim of universal superiority.

What you get

• The complete Hugging Face checkpoint: approximately 45.4 GB decimal / 42.3 GiB of weights, plus tokenizer, processor, remote model code and inference helpers. See release-manifest.json for exact uploaded sizes and hashes.
• An actual weight edit, not a system-prompt jailbreak: OBLITERATUS difference-of-means extraction from layer 22, control-PC denoise 0.2, followed by global projection of residual output weights and token embeddings.
• Vision weights and buffers preserved exactly; a synthetic red-square recognition test passed after a fresh reload. This is a smoke test, not a vision benchmark.
• 50/64 on our fixed capability sample, versus 51/64 for the unedited source loaded through the same NF4 build workflow. That small sample does not establish broad capability parity.
• Loading helpers for the hybrid Mamba/attention/MoE architecture, including its cache adapter and sparse expert inference path.

Comparison: 169 prompts, thinking off

All figures below count judge-labeled REFUSAL, not incorrect answers, partial answers or attack success. Lower refusal is not equivalent to higher answer quality.

BOLTBREAKER still refused nearly half of the general-harm bucket. Do not describe this release as zero-refusal or fully uncensored.

The cyber breakdown matters: BOLTBREAKER received 13 COMPLY / 14 PARTIAL / 3 REFUSAL, compared with the stock baseline's 1 COMPLY / 13 PARTIAL / 16 REFUSAL. Partial answers are counted separately from both compliance and refusal.

Benign answers were 41 COMPLY / 3 PARTIAL / 1 REFUSAL for BOLTBREAKER, versus 39 COMPLY / 2 PARTIAL / 3 REFUSAL / 1 UNCLEAR for the stock baseline.

What this comparison does and does not show

• The stock refusal baseline is Super-VL-120B, a different model. It is not an apples-to-apples refusal comparison against the original Bolt 74B.
• The 169 prompts came from StrongREJECT (42), HarmBench (42), XSTest (55), and custom cyber/pentest prompts (30). Category buckets are assigned separately: 94 harm, 45 benign, 30 cyber.
• Temperature 0, thinking disabled, maximum generation 420 tokens. BOLTBREAKER's server allowed 4,096 total prompt/output tokens.
• Responses were labeled by an automated language-model judge. Results are not independently human-adjudicated; judge bias is possible. The judge's identity and configuration are not specified in this public summary, which limits reproducibility.
• Stock rows were reused from an earlier run on the same prompts. Inference engines, precision and hardware differ.
• Sample sizes are small. Differences of a few answers are descriptive observations, not demonstrated statistical superiority.
• See evaluation/scores_baseline.json for the BOLTBREAKER and stock baseline label counts. Raw prompt/response transcripts are not included in this release.

Comparison against the actual Bolt source

During build validation, the unedited Bolt source was loaded through the same NF4 workflow and scored 51/64 on a fixed MMLU sample; the final edit scored 50/64. Eight short text checks and a fresh-process reload passed. A separate 32-prompt, 40-token prefix audit changed from 32 detected refusals to 1.

That short-prefix audit is not interchangeable with the later 169-prompt, model-judged evaluation. Use the broader results above when describing refusal behavior. The published NVIDIA BF16 model's full benchmark suite has not been rerun here.

Who this release is for

Try BOLTBREAKER if you have Spark-class memory, want a Hugging Face NF4 artifact with unmodified vision components, and care about the lower benign over-refusal observed in this sample.

This BOLTBREAKER checkpoint does not fit entirely in a 32 GiB RTX 5090. No speed advantage, quality advantage from precision, or universal superiority is claimed.

Loading and compatibility

This is a custom-code Transformers/bitsandbytes NF4 checkpoint, not GGUF. It is not validated as a drop-in model for Ollama, llama.cpp, stock vLLM, or every Transformers environment.

The validated runtime used ARM64 Python 3.12.3, CUDA PyTorch 2.14.0+cu130, Transformers 5.19.0, bitsandbytes 0.50.2, accelerate 1.15.0 and safetensors 0.8.0. See runtime-manifest.json and RUNTIME.txt. The required native/Triton dependencies make this an advanced deployment rather than a one-command install on arbitrary hardware.

Download the complete repository using the current hf CLI:

hf download BlackwellBoy/Boltbreaker-Nemotron-3.5-74B-A9B-NF4 --local-dir ./boltbreaker

In a compatible CUDA environment, run the supplied loader from its directory:

cd boltbreaker
python load_bolt.py . --prompt "What is 17 * 19? Reply with only the integer."

The loader uses CUDA device 0, reads saved NF4 metadata, verifies expected parameters and installs the model's cache/sparse-expert adapters. It loads repository Python code via trust_remote_code=True; review that code before running it. Keep all model shards and helper files together. Run programmatic inference inside torch.inference_mode().

The CLI smoke-test prompt is bounded to 128 tokens. The builder's separate HTTP wrapper was tested as a text-only, non-streaming service with a 4,096-token total limit and thinking off. Those are wrapper limits, not a newly established architectural context limit. Longer context and thinking-on quality were not validated for this derivative. Image support exists in the checkpoint; the evaluated HTTP wrapper did not accept images.

Provenance

• Source: NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16, revision a6a1e5030819a130a1ea08aadb1fe42f4cc86777.
• Extraction/projection tooling: OBLITERATUS, commit c968a8bc4007efd72b24174c1d35bfd542caff75, with the included Bolt-specific adapter.
• Build/evaluation date: 17 October 2026, Australia/Sydney.
• Release branding changes the repository name; checkpoint weights are the verified original build. No additional training or weight changes were made for this upload.
• Independent derivative by BlackwellBoy; not an official NVIDIA release.
• Source model card is retained as SOURCE_README.md; its claims describe the upstream model, not validation of this derivative.

License: OpenMDW 1.1. See LICENSE.OpenMDW-1.1 and retained upstream notices.