Math reasoning researchers
Study a specialized open-weight model and an explicit multi-candidate reasoning pipeline for difficult proof problems.
Independent tool overview
Nomos 1 at a glance
Nomos 1 is a 31-billion-parameter open-weight model specialized from Qwen3-30B-A3B-Thinking for mathematical problem solving and natural-language proof writing.
Visit the official Nomos 1 site ↗
- Product type
- Open-weight mathematical reasoning model
- Developer
- Nous Research with Hillclimb AI
- Base model
- Qwen3-30B-A3B-Thinking-2507
- Model size
- 31B parameters
- Model license
- Apache 2.0
- Harness license
- MIT
- Primary access
- Download and self-host
- Last reviewed
- August 29, 2026
Overview
What Nomos 1 is
Nomos 1 is a specialist mathematical reasoning model released by Nous Research in collaboration with Hillclimb AI. It is based on Qwen3-30B-A3B-Thinking-2507 and is intended for difficult problems that require a written proof rather than a short numeric answer.
The model is designed to work with the separate open-source Nomos reasoning harness. That harness generates many candidate solutions in parallel, scores them, consolidates similar conclusions, and uses pairwise comparisons to select a final submission.
Nous Research reports an 87/120 score on the 2025 Putnam problems with Nomos 1 plus the harness, compared with 24/120 for the base Qwen model under the same setup. That result describes a compute-intensive system evaluation, not the quality of a single ordinary chat response.
The weights are available on Hugging Face under Apache 2.0, while the reasoning harness is published on GitHub under MIT. Users can run the model through Transformers, vLLM, or SGLang, but the official examples assume substantial accelerator capacity and recommend eight-way tensor parallelism for the serving frameworks.
Nomos 1 is best treated as a research and experimentation release. Its proofs can still contain hidden gaps, unjustified steps, or confident errors, so important results should be checked by a qualified human or a formal proof system.
Use cases
Who Nomos 1 is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Olympiad and Putnam experimentation
Generate and compare candidate natural-language proofs for competition-style questions with expert evaluation.
Self-hosted model teams
Run math inference on controlled infrastructure through Transformers, vLLM, SGLang, or an OpenAI-compatible local endpoint.
Reasoning-system developers
Adapt the harness's parallel solving, self-critique, consolidation, and pairwise-selection workflow.
Model evaluation work
Compare the specialized checkpoint with its Qwen base model while holding the surrounding harness and grading process constant.
Open-model education projects
Explore generated proof drafts where instructors or experts can carefully verify and annotate every step.
Capabilities
Core Nomos 1 features
Math-specialized checkpoint
Fine-tunes a Qwen3 thinking model for mathematical problem solving and natural-language proof writing.
Open weights
Provides downloadable model files on Hugging Face under an Apache 2.0 license.
Nomos reasoning harness
Pairs the model with an MIT-licensed orchestration system for repeated solving, judging, and final selection.
Parallel candidate generation
The harness can launch multiple workers to create several possible solutions for each problem.
Self-scoring loop
Candidate submissions are scored on a seven-point scale until the target number of strong solutions or the time limit is reached.
Consolidation stage
The harness groups candidate solutions by conclusion and selects a group before final comparison.
Pairwise tournament
Remaining candidates are compared in a single-elimination selection process to produce the final proof.
Flexible serving
Official instructions cover Hugging Face Transformers, vLLM, and SGLang.
OpenAI-compatible endpoint
The harness can call a locally served model through a configurable OpenAI-style API base URL.
Published problem artifacts
The repository includes problems, prompts, runbooks, and generated submissions that help users inspect the evaluation setup.
Process
How the Nomos 1 workflow works
Step 1
Confirm the use case
Choose a proof-writing or difficult math task where an open model and substantial inference cost are justified.
Step 2
Review the licenses
Check the Apache 2.0 model terms, the MIT harness terms, and any obligations inherited from the base model or deployment stack.
Step 3
Provision inference hardware
Select accelerators and a model format appropriate for a 31B-parameter checkpoint; the official vLLM and SGLang examples use tensor parallelism across eight devices.
Step 4
Serve the model
Load Nomos 1 with Transformers or expose it through vLLM or SGLang at an OpenAI-compatible endpoint.
Step 5
Prepare problem files
Store each problem as Markdown and preserve all assumptions, definitions, and notation needed for a self-contained proof.
Step 6
Run a controlled baseline
Start with one model response and the recommended no-system-prompt setup before adding the more expensive harness.
Step 7
Use the reasoning harness
Configure the time limit, concurrency, scoring target, judge model, and prompts, then generate and compare multiple candidates.
Step 8
Verify the final proof
Check every inference independently, test boundary cases, and send consequential results to a human expert or formal verifier.
Step 9
Record reproducibility details
Save the checkpoint revision, prompts, sampling settings, harness commit, hardware, runtime, judge configuration, and all candidate outputs.
Cost
Nomos 1 pricing and free plan
Nous Research does not sell Nomos 1 as a subscription product. The model weights and reasoning harness are free to download under their respective open licenses, but users pay for the compute, storage, engineering, and any third-party hosting used to run them.
Model weights
Free to download
Nomos 1 is distributed on Hugging Face under Apache 2.0.
- 31B-parameter checkpoint
- No official Nomos subscription
- Users provide their own runtime and hardware
- Review all license obligations before commercial deployment
Reasoning harness
Free and open source
The orchestration code is available on GitHub under the MIT license.
- Parallel solving workflow
- Self-scoring and consolidation
- Pairwise candidate selection
- Compute usage rises with concurrency and time limit
Self-hosted inference
Infrastructure costs vary
Run the model on owned or rented accelerator infrastructure.
- Costs depend on hardware, precision, throughput, and runtime
- Official vLLM and SGLang examples use eight-way tensor parallelism
- The harness can make many concurrent requests
- Storage, monitoring, and engineering are separate costs
Third-party hosting
Provider-dependent
Use a compatible external host or managed endpoint if one supports the checkpoint.
- No first-party hosted price is published
- Availability varies by provider
- Check token pricing, data handling, and model revision
- Hugging Face currently lists no integrated inference provider for this model
Pricing checked . Check current pricing at the source ↗
Assessment
Nomos 1 strengths and limitations
Where it stands out
- Specialized specifically for mathematical reasoning and written proofs
- Open model weights make local inspection, adaptation, and reproducible evaluation possible
- Apache 2.0 licensing is comparatively permissive for a model release
- The separate MIT-licensed harness exposes how candidate generation and selection work
- Supports common inference stacks including Transformers, vLLM, and SGLang
- Published problems, prompts, runbooks, and submissions make the reported evaluation easier to inspect
- The reported Putnam result compares against the base model under the same harness
- OpenAI-compatible serving makes the harness adaptable to other models
What to consider
- This is a model checkpoint and research harness, not a polished end-user math application
- A 31B-parameter model requires considerably more hardware than small local models
- The official high-throughput serving examples assume eight-way tensor parallelism
- The full harness can generate many candidates and therefore consume far more compute than a single answer
- The 87/120 Putnam result depends on the harness, time budget, judging workflow, and human grading
- One benchmark does not establish reliability across all mathematical fields or proof styles
- Self-judging can repeat the solver's own blind spots and select a persuasive but incorrect proof
- Natural-language proofs may hide gaps that require expert or formal verification
- The model card provides limited detail about training data, post-training procedure, safety evaluation, and contamination controls
- Hugging Face does not list an integrated inference provider for the checkpoint
- Users are responsible for deployment security, privacy, monitoring, and cost controls
Compare
Nomos 1 alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Science
Gemini 3.1 Deep Think
Consider Gemini Deep Think for a managed frontier system aimed at advanced mathematical and scientific reasoning.
Explore Gemini 3.1 Deep Think →Consumer
Qwen3-Max-Thinking
Choose Qwen3-Max-Thinking when you want a broader flagship reasoning model rather than a self-hosted math specialist.
Explore Qwen3-Max-Thinking →Business Operations
DeepSeek
Consider DeepSeek for a broader open-model ecosystem with accessible reasoning models and hosted chat options.
Explore DeepSeek →Educators
Math Mentor
Choose Math Mentor for a simpler conversational tutoring experience instead of operating a research checkpoint and harness.
Explore Math Mentor →Questions
Nomos 1 FAQs
What is Nomos 1?
Nomos 1 is a 31B-parameter open-weight model from Nous Research and Hillclimb AI, specialized from Qwen3-30B-A3B-Thinking-2507 for mathematical problem solving and natural-language proof writing.
Is Nomos 1 free?
The model weights are free to download under Apache 2.0, and the reasoning harness is available under MIT. You still pay for hardware, cloud compute, storage, deployment work, and any third-party services.
Is Nomos 1 open source?
The downloadable model weights use Apache 2.0 and the harness code uses MIT. Because open-source terminology for AI models can be contested when full training data and recipes are not released, 'open-weight model with an open-source harness' is the most precise description.
How large is Nomos 1?
Hugging Face lists Nomos 1 at 31 billion parameters. It is based on the Qwen3-30B-A3B-Thinking mixture-of-experts model.
How do you run Nomos 1?
The official model card provides examples for Hugging Face Transformers, vLLM, and SGLang. The reasoning harness expects an OpenAI-compatible API endpoint and can point to a locally served instance.
What is the Nomos reasoning harness?
It is a separate orchestration system that creates candidate proofs in parallel, scores them, consolidates candidates by conclusion, and uses pairwise comparisons to select a final submission.
How did Nomos 1 score on Putnam 2025?
Nous Research reports 87/120 with the Nomos reasoning harness and human expert grading. The base Qwen checkpoint scored 24/120 under the same conditions.
Does the Putnam score represent one model response?
No. The reported system uses a multi-candidate harness with concurrency, repeated scoring, consolidation, pairwise selection, and a time budget, so it should not be interpreted as ordinary single-pass accuracy.
Can Nomos 1 proofs be trusted without checking?
No. A fluent proof can contain a subtle invalid step. Verify every conclusion independently and use qualified human review or formal proof tools for important work.
Bottom line
Our Nomos 1 verdict
Nomos 1 is a useful open research package for teams studying difficult natural-language math reasoning, especially because both the checkpoint and its multi-candidate harness are inspectable. Its headline Putnam result is interesting but comes from an expensive system-level workflow, not a single response. Use it when openness and experimentation outweigh ease of use, and budget for substantial inference plus rigorous proof verification.
Visit Nomos 1 website ↗Similar Science Tools
View all Science AI tools →
SciSpace BioMed Agent
SciSpace BioMed Agent - A specialized biomedical research co-scientist for multi-omics, clinical phenotypes, and lab protocol reasoning
TRIBE v2
TRIBE v2 - Meta's predictive foundation model that simulates human brain responses to sights, sounds, and language
Edison Analysis
Edison Analysis - Next-gen scientific analysis agent
Claude Science
Claude Science - Anthropic’s app for AI-powered scientific research
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.