NVIDIA-Nemotron-3.5-Bolt-E4B-4GB-BF16

NVIDIA-Nemotron-3.5-Bolt-E4B-4GB-BF16 is a small language model post-trained by NVIDIA from Gemma-4-E4B-it. It keeps the Gemma 4 E4B architecture and refreshes the weights and the chat template. The BF16 release is the full-precision checkpoint intended for customization, post-training, and evaluation. NeMo AutoModel provides full-parameter and Low-Rank Adaptation (LoRA) recipes that load the checkpoint through the native Gemma 4 path and apply the chat template shipped with the checkpoint.

Set up NeMo AutoModel with the latest container or follow the installation instructions.

Fine-Tune NVIDIA-Nemotron-3.5-Bolt-E4B-4GB-BF16

From the repository root, run the 8-GPU full-parameter recipe:

uv run automodel examples/vlm_finetune/nemotron_3_5_bolt/nemotron_3_5_bolt_e4b.yaml --nproc-per-node 8

Or run the 2-GPU LoRA recipe:

uv run automodel examples/vlm_finetune/nemotron_3_5_bolt/nemotron_3_5_bolt_e4b_peft.yaml --nproc-per-node 2

Choose a Workflow

Goal	Start Here
Full-parameter fine-tuning on MedPix-VQA	Use nemotron_3_5_bolt_e4b.yaml.
LoRA fine-tuning on MedPix-VQA	Use nemotron_3_5_bolt_e4b_peft.yaml.

Configuration

Setting	Full-parameter recipe	LoRA recipe
Hardware	8 NVIDIA H100 80GB GPUs	2 NVIDIA H100 80GB GPUs
Strategy	FSDP2; `tp_size=1`, `cp_size=1`	FSDP2; `tp_size=1`, `cp_size=1`
Batch	`local_batch_size=1`, `global_batch_size=8`	`local_batch_size=1`, `global_batch_size=8`
Attention	Eager	SDPA
Adapter	None	LoRA `dim=16`, `alpha=32` on the language model linear layers
Optimizer	AdamW, `lr=1e-5`	AdamW, `lr=2e-4`, cosine decay
Frozen modules	Vision and audio towers	Vision and audio towers, plus the base weights

Both recipes set dataset.chat_template and validation_dataset.chat_template to the chat template shipped with the checkpoint, so the training prompts match the template the model was post-trained with. The template exposes the same controls as the model card: enable_thinking switches reasoning on or off, and conversational_mode adds the spoken-conversation system instruction for concise, low-formatting replies.

Model Reference

Model Architecture

Property	Value
Developer	NVIDIA Corporation
Base model	`google/gemma-4-E4B-it`
Architecture type	Gemma 4
Network architecture	Gemma-4-E4B
Hugging Face architecture	`Gemma4ForConditionalGeneration`
Total parameters	8B (4B effective)
Precision	BF16
Context length	Input up to 64K tokens; output up to 64K tokens
Inputs and outputs	Text in, text out
Reasoning mode	Chat-template controls (`enable_thinking=True/False`) and a runtime thinking budget
Conversational mode	Chat-template control (`conversational_mode=True/False`)
Recommended sampling	Temperature 1.0, top_p 0.95 in both thinking modes
Best for	Local voice and device assistants, interactive personas, and embedded interfaces
License	OpenMDW-1.1
Checkpoint	`nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB-BF16`

Training Methodology

The model was post-trained from Gemma-4-E4B-it with multi-environment reinforcement learning using Group Relative Policy Optimization (GRPO) across instruction following, multi-step tool use, multi-turn conversations, and structured-output environments, on an asynchronous RL architecture that decouples training from inference. The reinforcement learning used NeMo RL and NeMo Gym.

Prompting Guidelines

The model card recommends the following for conversational and agentic applications:

Store persistent behavior constraints and operational guidelines in system prompts, with explicit headers for roles, tasks, tool configs, and formatting rules.
Write unambiguous, verifiable instructions and use explicit constraints such as “limit responses to 1-2 sentences” instead of vague requests.
Adhere strictly to standard system/user/assistant/tool turn sequences and avoid synthetic outputs in assistant turns.
Set conversational_mode=True for concise, low-formatting outputs suited to spoken interaction.

Available Models

nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB-BF16 (HF BF16)
nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB (Q4_K_M GGUF for inference)