Modelpublic

NexusProjectsAI/Nemotron-3-Nano-30B-A3B-Nexus-Agents-GGUF

sourceHugging Faceotherupdated 4mo agoView on Hugging Face

0likes183downloads

Model Card

Nemotron-3-Nano-30B-A3B — Nexus Agents (GGUF)

A LoRA fine-tune of nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B (hybrid Mamba/Transformer MoE, ~3B active) specialized for the Nexus Projects agent stack.

Links: the exact training + verification data → Nexus-Agents-ToolCalling · the tool that generated the data, trained, quantized, and evaluated this model → Nexus Training Studio · the app these agents power → Nexus Projects client

—Setup interview — infers industry/platforms/objectives from a free-text idea ("I want to sell lemonade" → Food & Beverage), asks instead of guessing when the input is vague or ambiguous ("a lemonade stand" → asks Food & Beverage or Retail?), and looks libraries up on the internet (pub.dev/GitHub) to pick current ones before finishing.
—Discovery — builds a user-story tree (As a … I want … so that …).
—Task generation — turns setup + stories into concrete, stack-specific tasks with acceptance criteria and verification commands.

How it was trained

Two LoRA stages (rank 16/scale 32, attention + Mamba mixer + always-on shared_experts MLP — never attention-only), on schema-verified synthetic tool-calling conversations carrying the agents' real tool schemas — train == serve. Both corpora are published in the dataset repo:

Stage	Dataset config	Rows	What it taught
1 — full LoRA	`stage1`	60,185	the three core skills
2 — additive recovery	`stage2-recovery`	16,015	error-recovery discipline (resume a broken board state, re-ask nothing, finalize), with ~50% stage-1 replay

Results — base vs fine-tuned

Behavioral interview eval (27 end-to-end agent scenarios, greedy decoding)

Each scenario is a full multi-turn setup interview driven against the served GGUF with the agents' real tool schemas. A case passes only if the model covers every required field, never re-asks an answered question, and cleanly finalizes. Full transcripts for every case (base + fine-tuned) are published in the dataset repo's `verification/` folder.

Metric	Base (no LoRA)	Fine-tuned
Scenarios passed	13 / 27	27 / 27
Asks each question at most once	52%	100%
Completes (finalizes) the interview	92%	100%
Avg redundant re-asks per interview	4.92	0.89

Tool-call accuracy (BFCL-style, 150 held-out calls)

Metric	Base (no LoRA)	Fine-tuned	Lift
Function-name exact	35.3%	95.3%	+60.0
Argument-keys match	32.7%	95.3%	+62.6
Arguments exact	12.0%	64.0%	+52.0

The base model only emitted a parseable tool call 109/150 times; the fine-tune did so 147/150. In short, fine-tuning ~tripled tool-use accuracy on the Nexus tool set.

Quantizations

All quants are imatrix-calibrated.

File	Bits	Size
`Nemotron-3-Nano-30B-A3B-Nexus-Agents-Q8_0.gguf`	8-bit	31.3 GB
`Nemotron-3-Nano-30B-A3B-Nexus-Agents-Q6_K.gguf`	6-bit	31.2 GB
`Nemotron-3-Nano-30B-A3B-Nexus-Agents-Q4_K_M.gguf`	4-bit	22.8 GB

imatrix.dat (the calibration importance matrix) is included for re-quantizing.

⚠️ Serving requirements

1.Disable thinking. Nemotron-3-Nano is a reasoning model; with thinking on it reasons in prose instead of calling tools. Serve with enable_thinking=false (renders <think></think>). It is also prompt-sensitive — use the agent's real system prompt.
2.Tool-call format is Nemotron's <tool_call><function=NAME><parameter=key>value</parameter>…</function></tool_call> — parse that (not JSON) on the client.

License

Inherits the NVIDIA Open Model License from the base model.