Open Source

If you've been living under a rock, Nvidia shipped Nemotron 4 340B with fully open weights under a commercial-friendly license. Spoiler: it's not for chat. It's for cooking training data for other models.

The Setup

Nemotron 4 ships as a base, an instruct, and a reward model — the full pipeline. The play here is using it to generate millions of synthetic SFT pairs to fine-tune a smaller model that actually runs in production.

# Hosted NIM endpoint — fastest way in
from openai import OpenAI

client = OpenAI(
    base_url="https://integrate.api.nvidia.com/v1",
    api_key="nvapi-...",
)

resp = client.chat.completions.create(
    model="nvidia/nemotron-4-340b-instruct",
    messages=[{"role": "user", "content": "generate 5 SQL training pairs"}],
    temperature=0.5,
)
print(resp.choices[0].message.content)

The Money Pattern

The reward model is the real sleeper. Pair instruct with reward and you've got a self-rating loop that ranks its own outputs. Plug into a Pydantic schema and you're filtering garbage data automatically.

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("nvidia/Nemotron-4-340B-Instruct")
model = AutoModelForCausalLM.from_pretrained(
    "nvidia/Nemotron-4-340B-Instruct",
    device_map="auto",
    torch_dtype="bfloat16",
)

# pair with reward model to filter low-quality generations
# nvidia/Nemotron-4-340B-Reward returns a scalar per response

The Catch

340B parameters means 8xH100s minimum to host. That's a $200k box. You're not running this on your M4 Mac, no matter how cute the new Pro chips are. The realistic play is hosted NIM until your data is cooked, then move on.

The Verdict

Nemotron 4 isn't trying to win chatbot benchmarks — it's trying to be the upstream model that trains every other model. For anyone shipping a domain-specific fine-tune, this is the cheapest way to print training data legally. Underrated drop.