MC7ever/darwin-nemotronh-30b-a3b-v0.1

Darwin NemotronH v0.1

A Darwin MRI-Trust evolutionary merge of 5 NemotronH 30B-A3B models + ToolCall LoRA, quantized from GGUF Q4KM and dequantized to float16.

Architecture

• —Model type: NemotronH (hybrid SSM/MoE/Attention)
• —Hidden size: 2688
• —Layers: 52 (26 Mamba + 20 MoE + 6 Attention)
• —MoE experts: 128 (6 active)
• —Attention heads: 32 (2 KV heads)
• —Max context: 262144 (256k)

Merge Components

1. 1.nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16
2. 2.thoughtworks/Nemotron-3.5-30B-A3B-Antislop-FTPO
3. 3.thoughtworks/Nemotron-3.5-30B-A3B-Antislop-FTPO-V2
4. 4.vcruz305/DeepNemotron-3.5-Lightning-Beta-BF16
5. 5.brazley/Brazley-Nemotron-3.5-Lightning-Omni

• —PursuitOfDataScience/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-ToolCall-LoRA

Merge Strategy

Darwin MRI-Trust Fusion with CMA-ES evolutionary genome search. Sequential pairwise merges across the model population.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "MC7ever/darwin-nemotronh-30b-a3b-v0.1",
    torch_dtype=torch.float16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("MC7ever/darwin-nemotronh-30b-a3b-v0.1")

Quantized Versions

GGUF quantized versions available (Q4KM, Q8_0).

License

Based on NVIDIA Nemotron models - check original model licenses.