Description
I am encountering a RuntimeError: Unsloth: No config file found when attempting to load unsloth/Nemotron-3-Nano-30B-A3B using FastLanguageModel.from_pretrained on a DGX Spark environment.

While manual download via hf_hub_download succeeds for the -Base version, the automated loader fails to identify the model configuration. This seems related to the internal mapping logic for the Nemotron-3 (Mamba-2 Hybrid) architecture in recent unsloth_zoo versions.

Environment

• Hardware: DGX Spark (NVIDIA GB10 / Blackwell)
• CUDA: 13.0
• Unsloth version: 2026.1.4
• Unsloth_zoo version: 2026.1.4
• Transformers version: 4.57.6
• Torch version: 2.9.0+cu130
• TRL version: 0.24.0

Reproduction Code

from unsloth import FastLanguageModel
import torch

max_seq_length = 2048
# Manual check: unsloth/Nemotron-3-Nano-30B-A3B exists and is downloadable
MODEL_NAME = "unsloth/Nemotron-3-Nano-30B-A3B" 

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = MODEL_NAME,
    max_seq_length = max_seq_length,
    load_in_4bit = True,
    token = "YOUR_HF_TOKEN", 
)

Traceback (most recent call last):
  File "/workspace/train.py", line 14, in <module>
    model, tokenizer = FastLanguageModel.from_pretrained(
                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/unsloth/models/loader.py", line 368, in from_pretrained
    model_types = get_transformers_model_type(
                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/unsloth_zoo/hf_utils.py", line 112, in get_transformers_model_type
    raise RuntimeError(
RuntimeError: Unsloth: No config file found - are you sure the `model_name` is correct?
If you're using a model on your local device, confirm if the folder location exists.
If you're using a HuggingFace online model, check if it exists.

Analysis of Provided Source Code

• loader.py logic: The FastLanguageModel.from_pretrained function fails during the get_transformers_model_type call. On DGX Spark (GB10), the loader defaults to native Unsloth inference as fast_inference=True is noted as broken for this hardware in the source.
• peft_utils.py logic: Nemotron-3's hybrid Mamba-2 architecture layers (like mamba) are explicitly listed in SKIP_QUANTIZATION_MODULES. If the MODEL_NAME string doesn't precisely match the registry (e.g., missing -Base), the mapping fails to find the config.json.
• Manual Verification: * hf_hub_download works for unsloth/Nemotron-3-Nano-30B-A3B.

Additional Context

• This issue seems similar to discussions in [Bug] Unsloth: No config file found - are you sure the model_name is correct?总是报错，模型明明下载好 #3408, No config file found - are you sure the model_name is correct? If you're using a model on your local device, confirm if the folder location exists. If you're using a HuggingFace online model, check if it exists. #3479, and [Bug] RuntimeError: Unsloth: No config file found - are you sure the model_name is correct? for gpt-oss colab notebook #3430.