---

language:

- en

license: other

license_name: nvidia-openmdw-and-dataset-other

license_link: https://openmdw.ai/license/1-1/

library_name: gguf

pipeline_tag: text-generation

base_model: vcruz305/DeepNemotron-3.5-Lightning

base_model_relation: quantized

datasets:

- r0b0tlab/deepseek-v4-pro-0813-agentic

tags:

- gguf

- llama.cpp

- nemotron

- moe

- quantized

- agentic

---

DeepNemotron-3.5-Lightning-GGUF

llama.cpp K-quants of the merged M5 LoRA on NVIDIA Nemotron 3.5 Lightning 30B-A3B.

Adapter: vcruz305/DeepNemotron-3.5-Lightning

BF16: vcruz305/DeepNemotron-3.5-Lightning-BF16

MXFP4: vcruz305/DeepNemotron-3.5-Lightning-MXFP4

NVFP4: vcruz305/DeepNemotron-3.5-Lightning-NVFP4

llama-server -hf vcruz305/DeepNemotron-3.5-Lightning-GGUF:Q4_K_M

K-quants on this repo: Q4_K_M, Q5_K_M, Q2_K, Q8_0, Q6_K, Q3_K_M.

Train set

SFT used r0b0tlab/deepseek-v4-pro-0813-agentic config sft_openai, train split (19,072 rows), max seq 512.

DS4 is a verifiable-first agentic set taught by DeepSeek-V4-Pro 0813 (deepseek-v4-pro-0813, official API, thinking on). Each row is admitted after a deterministic programmatic verifier. Published split is 19,072 / 1,070 / 1,065 (train / val / test), seed-grouped, 13 task families:

| Family | Rows (all splits) |

|---|---:|

| tool_call_schema_adherence | 5,541 |

| verified_constraint_instruction | 4,210 |

| structured_outputs | 1,951 |

| stateful_dialogue | 1,856 |

| verifiable_math | 1,812 |

| science_verifiable | 1,748 |

| multilingual_multi_turn | 1,649 |

| planning_decomposition | 761 |

| grounded_long_context | 551 |

| delegation_orchestration | 397 |

| grounded_citations | 361 |

| memory_context_management | 194 |

| code_execution | 176 unique solutions |

Train (MEASURED)

• Modal A100-80, Unsloth 2026.8.18, torch 2.7.1+cu118
• Q4′ LoRA r=16 α=32 · attn q/k/v/o + Mamba in_proj + shared-expert Linears
• 19,072 / 19,072 steps, epoch 1.0, train_loss 0.151
• Metered $22.72 on Modal free credits; $0 billed

License

NVIDIA Nemotron 3.5 Lightning (OpenMDW) and dataset license other (DeepSeek API-adjacent). Use is under the base card and the dataset card.