NVIDIA Drops New Open Code Reasoning Models - Fast Code Thinking Just Got Cheaper

• NVIDIA launched 7B 14B and 32B Open Code Reasoning models under Apache 2.0
• Beats O3 Mini and O1 on LiveCodeBench
• Works with vLLM llama.cpp Transformers TGI and more
• Community reactions are mixed but curious

NVIDIA just tossed out three new models in their Nemotron family - 7B 14B and 32B - all open and Apache 2.0 licensed. That last part matters cause you can actually build stuff with them with no drama.

Here’s the kicker. These models pull off better scores than O3 Mini and O1 (low) on LiveCodeBench. According to NVIDIA they’re also 30% more efficient when it comes to token use thanks to their OCR dataset.

They also play nice with llama.cpp vLLM Transformers TGI and probably a few more by the time you finish reading this.

You can check them out right now:

• Nemotron 14B
• Nemotron 7B
• Nemotron 32B IO

People on Local Llama subreddit so far are hyped but cautious.

One user on reddit wrote: “The 32B almost benchmarks as high as R1 but I don’t trust benchmarks anymore… so I suppose I’ll wait for vram warriors to test it out.”

@DeProgrammer99 pointed out the naming doesn’t help: “The fact that they call their own model 'OCR-Qwen' doesn’t help the readability.”

@SomeOddCodeGuy was more upbeat saying: “I’ve always liked NVidia’s models... each one in the Nemotron family has been great for productivity.”

Meanwhile @LocoMod confirmed it already runs in llama.cpp and dropped a working 👉 GGUF: 👉 Link

But not everyone’s sold. One user said: “Tried Olympic Coder 7b and 32b and it felt worse than Qwen 2.5. Hoping I’m wrong.”

And there’s still no instruct versions out yet which left @DinoAmino saying: “They print benchmarks for both base and instruct models. But I don’t see any instruct models :(”

So... Worth Trying?

If you’ve got VRAM to spare and like tinkering with code models this could be a solid option. If not maybe wait till the instruct versions drop.

The field’s moving fast. Mistral dropped a new Medium 3 model the same day which aims to offer solid results at lower cost for enterprise stuff like code and function calling.

Time will tell which one really wins out.