This Open-Source AI Just Beat Most Humans at Math's Most Brutal Competition (And It's Free)
Here's the thing that'll make you spit out your coffee: an open-source AI just scored 87 out of 120 on the 2025 Putnam exam. If you've never heard of the Putnam, imagine the mathematical equivalent of Navy SEAL training crossed with a Rubik's cube designed by sadists.
Nous Research dropped Nomos 1 on December 9th, and honestly? This feels like a watershed moment that nobody saw coming.
The Putnam Mathematical Competition is where math prodigies go to get humbled. We're talking about problems so brutal that scoring above 50 puts you in rarified air. Most participants - and these are students from MIT, Harvard, Caltech - walk away with single-digit scores. If they're lucky.
<> The model scored only 24/120 under identical conditions without the specialized reasoning harness, jumping to 87/120 with it - a performance leap that defies typical AI scaling patterns./>
But here's where it gets interesting. Nomos 1 isn't some mystical breakthrough model. It's a fine-tuned version of Qwen/Qwen3-30B-A3B-Thinking-2507, developed with Hillclimb AI. The magic sauce? Something called the Nomos Reasoning Harness that uses AI "workers" to solve problems, self-critique, then duke it out tournament-style.
Without that harness? The base model scored a measly 24 points. With it? Suddenly we're in elite human territory.
What Nobody Is Talking About
Everyone's celebrating the Putnam score, but they're missing the real story: this thing runs on 8 GPUs and costs you exactly zero dollars. While OpenAI and Google are charging premium prices for their math reasoning, Nous just handed everyone the keys to the kingdom.
You can literally fire this up right now:
- Download from NousResearch/nomos-1 on Hugging Face
- Deploy via SGLang or vLLM
- No system prompt needed
- Start solving graduate-level proofs
The technical specs are surprisingly modest for what it achieves. We're talking about a 30B-parameter model that you can run with device_map="auto" and trust_remote_code=True. That's it. No PhD in distributed systems required.
This is Nous Research flexing. They already made waves with Psyche back in May - a 40B-parameter beast that could train on consumer hardware like RTX 3090s. Now they're proving that open-source doesn't mean second-rate.
The early adoption metrics tell the story: 132 downloads in the first month. That might sound small, but in the world of 30B models requiring 8-GPU setups, those are serious players kicking the tires.
The Uncomfortable Truth
Here's my hot take: the reasoning harness is doing heavy lifting here. That jump from 24 to 87 points isn't just model improvement - it's architectural wizardry. The tournament-style self-critique system essentially gives the AI multiple attempts and internal debate.
Is that cheating? Absolutely not. It's brilliant engineering. But it does mean the "model" that scored second place is really a system - base model plus reasoning scaffold plus computational orchestration.
The implications are staggering. Every startup building math tutoring apps, every researcher needing proof verification, every developer who's been waiting for accessible mathematical reasoning - they just got handed a Christmas present in December.
Nous Research isn't just releasing models anymore. They're democratizing mathematical intelligence. And frankly, the big tech companies should be sweating.
Because when elite performance becomes free and open-source, the whole game changes.
