Nemotron-3-Nano-Omni-30B-A3B-Reasoning — APEX GGUF

APEX (Adaptive Precision for EXpert Models) quantizations of nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16.

Brought to you by the LocalAI team | APEX Project | Technical Report

Available Files

File	Profile	Size	Best For
Nemotron-3-Nano-Omni-30B-A3B-Reasoning-APEX-I-Balanced.gguf	I-Balanced	26 GB	Best overall quality/size ratio
Nemotron-3-Nano-Omni-30B-A3B-Reasoning-APEX-Balanced.gguf	Balanced	26 GB	General purpose
Nemotron-3-Nano-Omni-30B-A3B-Reasoning-APEX-I-Quality.gguf	I-Quality	22 GB	Highest quality with imatrix
Nemotron-3-Nano-Omni-30B-A3B-Reasoning-APEX-Quality.gguf	Quality	22 GB	Highest quality standard
Nemotron-3-Nano-Omni-30B-A3B-Reasoning-APEX-I-Compact.gguf	I-Compact	19 GB	Consumer GPUs, best quality/size
Nemotron-3-Nano-Omni-30B-A3B-Reasoning-APEX-Compact.gguf	Compact	19 GB	Consumer GPUs
Nemotron-3-Nano-Omni-30B-A3B-Reasoning-APEX-I-Mini.gguf	I-Mini	18 GB	Smallest "safe" tier
Nemotron-3-Nano-Omni-30B-A3B-Reasoning-APEX-I-Nano.gguf	I-Nano	17 GB	Experimental — IQ2_XXS mid-layer experts
Nemotron-3-Nano-Omni-30B-A3B-Reasoning-F16.gguf	F16 reference	59 GB	Full-precision reference (text-only)
mmproj.gguf	Vision+audio projector	~1.6 GB	Required for image and audio understanding

What is APEX?

APEX is a quantization strategy for Mixture-of-Experts (MoE) models. It classifies tensors by role (routed expert, shared expert, attention) and applies a layer-wise precision gradient — edge layers get higher precision, middle layers get more aggressive compression. I-variants use diverse imatrix calibration (chat, code, reasoning, tool-calling, agentic traces, Wikipedia).

The key insight: in MoE models, expert FFN tensors make up the bulk of model weight but only 6/128 experts activate per token. APEX compresses middle-layer experts more aggressively while preserving edge layers (first/last 5) and keeping attention, SSM/Mamba, and shared expert tensors at higher precision.

See the APEX project for full details, technical report, and scripts.

Nano (experimental tier)

The APEX Nano tier pushes mid-layer routed experts to IQ2_XXS (2.06 bpw), near-edge to IQ2_S, edges to Q3_K, with shared experts kept at Q5_K. About 5% smaller than Mini with modest quality cost — viable only on MoE thanks to sparse per-token expert activation. Requires imatrix.

Benchmarks pending. Feedback welcome.

Multimodal Support

This is the Omni variant — supports text + vision + audio inputs. The included mmproj.gguf (sourced from unsloth) provides:

• Vision: RADIO ViT encoder (1280-dim)
• Audio: Parakeet encoder (1024-dim, 24 layers)

Pass --mmproj mmproj.gguf to llama.cpp / LocalAI to enable multimodal inference. Note: llama.cpp's audio output is not yet supported in mtmd — audio input only.

Architecture

• Outer model: NemotronH_Nano_Omni_Reasoning_V3 (multimodal wrapper)
• Inner LLM: NemotronH (NemotronHForCausalLM) — same as Nemotron-3-Nano-30B-A3B
• Layers: 52 (23 Mamba-2, 23 MoE, 6 attention) per pattern MEMEM*EMEMEM*EMEMEM*EMEMEM*EMEMEM*EMEMEMEM*EMEMEMEME
• Experts: 128 routed + 1 shared (6 active per token)
• Total Parameters: 30B (LLM only) + RADIO + Parakeet
• Active Parameters: ~3.5B per token
• Hidden size: 2688
• Context: 262,144 tokens
• APEX Config: 5+5 symmetric edge gradient across 52 layers
• Calibration: v1.3 diverse dataset (chat, code, reasoning, multilingual, tool-calling, Wikipedia)

Run with LocalAI

local-ai run mudler/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-APEX-GGUF@Nemotron-3-Nano-Omni-30B-A3B-Reasoning-APEX-I-Balanced.gguf

Credits

• Base model: NVIDIA Nemotron team
• Vision+audio mmproj: unsloth
• APEX quantization: LocalAI team
• Built on llama.cpp (with PR #22481 — Nemotron Nano 3 Omni convert support)