license: other
license_name: openmdw-1.1
license_link: https://openmdw.ai/license/1-1/
language:

• en
  tags:
• nemotron-h
• darkstar
• abliteration
• nvfp4
• modelopt
• hybrid-mamba-moe
  pipeline_tag: text-generation
  base_model: HangGlidersRule/Darkstar-Nemotron-3.5-Lightning-30B-A3B-Abliterated-BF16
  base_model_relation: quantized
  quantization: nvidia-modelopt
  extra_gated_heading: Darkstar Nemotron-3.5-Lightning 30B-A3B Abliterated ModelOpt NVFP4

Darkstar-Nemotron-3.5-Lightning-30B-A3B-Abliterated-ModelOpt-W4A16-NVFP4

Safety notice: This checkpoint is an edited (abliterated) derivative
served in NVFP4: the refusal direction measured at layer 34 was
deliberately projected out of 3,126 residual-writing tensors, then the edited
BF16 artifact was quantized with NVIDIA TensorRT Model Optimizer to a mixed
W4A16-NVFP4 layout. It will tend to comply with harmful requests. Released for
red-teaming and alignment research only. Also see Safety.

A single-GPU-friendly (≈22 GB), modelopt-quantized derivative of NVIDIA's
Nemotron-3.5-Lightning-30B-A3B-BF16:
hybrid Mamba2 + MoE + sparse attention, 52 layers, 262,144-token context,
OpenMDW-1.1 license. This is the fourth product in the Darkstar family matrix.

Product family

Quantization contract (reproducible)

• Source for the edit: nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16
  @ d468880b6ad3c6e0d21377ce7242adaea4cc884d
• Abliteration: identical to the Abliterated-BF16 product (layer 34, 3,126
  targets, max residual leakage 0.000160, MTP intact, 320/320 chat-templated)
• Quantization: NVIDIA Model Optimizer 0.46.0rc2 (43fd41a), recipe
  w4a16_nvfp4_mse-fp8_attn-kv_bf16_nemotron_h.yaml
• Calibration: cnn_dailymail 512 + Nemotron-Post-Training-Dataset-v2 512,
  sequence length 2048, seed 1234, batch 1, KV cache quantization disabled (BF16)
• Protected BF16: lm_head, Mamba/SSM (conv1d, in_proj, out_proj, A_log, D,
  dt_bias), norms, embeddings, MTP head
• Quantized: routed + shared expert up/down projections (5,934 modules),
  W4A16-NVFP4 group 16
• Artifact: 3 shards, 22 GB, quantization_config → vLLM modelopt_mixed
• Recipe: recipes/nemotron-3.5-lightning/darkstar-nemotron-3.5-lightning-30b-a3b-abliterated-modelopt-nvfp4.yaml
  in HangGlidersRule/model-forge

Measured quality (final servable result)

• GPQA Diamond: 141/198 = 71.2% (Wilson 95% 64.5–77.1; 1 unparseable, 2 TRUNC, 0 errors),
  evaluated with llm-inference-bench gpqa-diamond (chat template + thinking ON, temp 0),
  served MTP10 + --reasoning-parser nemotron_v3, single RTX PRO 6000 Blackwell, BF16 KV.
• Behavior gate: 200/200 harmful compliance, 0/83 safe over-refusals, 0 errors.
• Throughput (weighted 4K/16K/48K = 0.6/0.3/0.1): MTP10 = 554.7 tok/s (4K 571.2, 16K 546.3, 48K 480.7).
• NVIDIA publishes GPQA Diamond 75.44 (BF16) / 75.57 (NVFP4) on the same task; the delta is
  serving-stack config (their vLLM 0.26 + FP8 KV + TP2 + temp 1.0 averaged over 8 repeats), not
  abliteration or quantization damage. Full protocol: models/nemotron-3.5-lightning-r1/gpqa-protocol.md.

License

OpenMDW-1.1 (same as upstream). See https://openmdw.ai/license/1-1/. Retain all
NVIDIA copyright/attribution/notice lines in distributions of this derivative.

Publication

This checkpoint is public on Hugging Face at the pinned milestone tag
darkstar-nemotron-3.5-lightning-v1.0.0. Weights are hash-verified (sha256
manifest in the source repo) and serve with vLLM (the measured shipping
config — MTP10):

vllm serve HangGlidersRule/Darkstar-Nemotron-3.5-Lightning-30B-A3B-Abliterated-ModelOpt-W4A16-NVFP4 \
  --max-model-len 131072 --kv-cache-dtype bfloat16 --reasoning-parser nemotron_v3 \
  --speculative-config '{"method":"mtp","num_speculative_tokens":10}'

Safety

This model intentionally has a reduced refusal response. Do not deploy in
user-facing assistant roles without alignment hardening and content filtering.
It is intended for researchers studying refusal behavior, ablation, and
alignment techniques.