What is Nomos 1?
Nomos 1 is a maths specialist, not a chat assistant. Nous Research post-trained Qwen's 30B mixture-of-experts thinking model on competition problems, and meant it to run inside the Nomos Reasoning Harness: an open-source loop in which many copies attempt each problem, grade their own proofs out of 7, and then pick one answer per problem in a pairwise tournament. Nous says most of the gain comes from post-training, not from the harness.
It is still Nous's current maths model, and nothing has replaced it. Nous publishes only the weights and the harness: there is no hosted API, app or paid plan.
Who is Nomos 1 best for?
Nomos 1 is best for competition-style maths and proof-writing on your own hardware, and for researchers who want an open maths setup they can rerun and check.
Competition maths and proofs
It targets Putnam and olympiad-style problems and writes full proofs in plain language, not just final answers.
Maths reasoning research
The weights, the harness, its prompts and the graded Putnam answers are all public, so the result can be rerun and compared.
Teams with their own GPUs
It is free to run under Apache 2.0 with Transformers, vLLM or SGLang, and Nous's own examples serve it across 8 GPUs.
Private or offline work
Because it runs only on hardware you control, problems and draft proofs never leave your servers.
What's included with Nomos 1?
An open reasoning harness that runs parallel attempts, grades them and picks one answer by tournament, plus a long context and public graded results.
- Nomos Reasoning Harness
- An MIT-licensed Python loop that runs many attempts in parallel, scores each out of 7, and keeps going until a time limit or a target number of perfect scores.
- Tournament final pick
- Near the time limit it groups answers by their conclusion, then runs a single-elimination bracket to choose one submission per problem.
- Any OpenAI-compatible server
- The harness talks to vLLM, SGLang or any server with an OpenAI-style API, and Nomos 1 acts as both solver and judge by default.
- Long reasoning window
- It keeps its base model's 262,144-token context, and Nous recommends running it without a system prompt.
- Public graded answers
- Nous published the exact Putnam answers its grader marked, plus the runbooks that produced them, so anyone can check the score.
How does Nomos 1 work?
Serve the model
Download the weights from Hugging Face and serve them with vLLM or SGLang behind an OpenAI-compatible endpoint.
Write the problems
Put each problem in its own Markdown file inside one folder.
Run the harness
Start solve_agent.py on that folder with a time limit; it defaults to 3 hours and 32 parallel requests.
Read the answers
Each final submission lands in its own Markdown file, with the problem text followed by the chosen proof.
How much does Nomos 1 cost?
Nomos 1 is free: the weights are an Apache 2.0 download, and the only cost is the hardware you run it on.
- About 122 GB of full-precision weights on Hugging Face
- Reasoning harness on GitHub under the MIT licence
- No hosted API, app or paid plan from Nous
Running Nomos 1 costs only your own hardware or rented GPUs. Hugging Face lists no inference provider for it, and it is not in the Nous Portal model catalog. Smaller quantized versions (GGUF, MLX, AWQ and FP8) exist on Hugging Face, but they come from the community, not from Nous. Full pricing on huggingface.co. We checked these prices on .
How do readers rate Nomos 1?
No one has rated Nomos 1 yet. Be the first.
Sign in to rate. One rating per account.
Nomos 1 pros and cons
It posts a strong Putnam score for a free 30B model, but it needs serious hardware and Nous's harness running for hours to get there.
Pros
- Scores 87 of 120 on the Putnam with Nous's harness, a result Nous says would have placed second among 3,988 human entrants the year before.
- The gain comes from training, not tooling: its Qwen base model scored 24 of 120 in the same harness, and Reddit users were struck that a 30B model did this within the contest's own three-hour sessions.
- Fully open and checkable: Apache 2.0 weights, an MIT harness, and the graded answers and runbooks are all public.
Cons
- Heavy to run: Nous ships only full-precision weights of about 122 GB and serves them across 8 GPUs in its examples, and Reddit users waiting to run it locally asked for smaller builds.
- No hosted option: there is no API, app or chat window from Nous, so trying it means setting up your own server.
- The headline score depends on the harness running many parallel attempts for hours, and Nous has published no scores for the model on its own.
- Narrow by design: it is tuned for maths problems and proofs, and Nous publishes no results for coding, chat or general tasks.
Is Nomos 1 open source?
Yes. Nomos 1's code is public under the Apache 2.0 licence, so you can run it on your own server.
What are the best Nomos 1 alternatives?
Nomos 1 is an open source alternative to Gemini 3.1 Deep Think. DeepSeek-V3.2-Speciale and NousCoder-14B do the same job and are open source too.
Gemini 3.1 Deep ThinkPaidChoose it when you want olympiad-level maths without running any hardware; Google reports 81.5% on IMO 2025, and it needs Google AI Ultra from $99.99 a month.
DeepSeek-V3.2-SpecialeOpen SourceDeepSeek's high-compute V3.2 variant for hard math and coding problems, released as open weights.
NousCoder-14BOpen SourceAn Apache-licensed 14B model tuned for competitive programming.
Leanstral 1.5Open SourceChoose it when you need proofs a computer can verify: it writes Lean 4 rather than plain-language proofs, and Mistral offers it on a free API as well as open Apache 2.0 weights.
Falcon H1R 7BFreeChoose it when you need a maths reasoning model that fits on one GPU; its weights are about 15 GB against 122 GB for Nomos 1, though TII tests it on answer-based contests such as AIME rather than written proofs.
Questions about Nomos 1
What is Nomos 1?
Nomos 1 is Nous Research's free, open-weights maths model, a fine-tune of Qwen3-30B-A3B-Thinking-2507 built with Hillclimb AI to solve competition problems and write proofs in plain language.
Who is Nomos 1 for?
Nomos 1 is best for competition-style maths and proof-writing on your own hardware, and for researchers who want an open maths setup they can rerun and check.
Is Nomos 1 free?
Yes. Nomos 1 is a free download under the Apache 2.0 licence, which allows commercial use, and its reasoning harness is MIT-licensed. Nous sells no hosted version, so the only cost is the hardware or rented GPUs you run it on.
Is Nomos 1 open source?
Yes. Nomos 1's code is public under the Apache 2.0 licence, so you can run it on your own server.
What are the best Nomos 1 alternatives?
The closest alternatives on Recaply are Gemini 3.1 Deep Think, DeepSeek-V3.2-Speciale, NousCoder-14B, Leanstral 1.5 and Falcon H1R 7B.
Where can I try Nomos 1?
On its official website, huggingface.co.
Can I use Nomos 1 without my own GPU server?
Not through Nous. Hugging Face lists no inference provider for Nomos 1 and the Nous Portal does not offer it, so you run it yourself. Community quantized builds on Hugging Face, such as GGUF and MLX versions, need less memory but are not made by Nous.
Do I need the Nomos harness to use Nomos 1?
No. Nomos 1 answers like any chat model, and Nous suggests leaving out the system prompt. But the 87/120 Putnam score comes from the harness, which runs many attempts in parallel and picks the best, and Nous has not published scores for the model alone.
Is Nomos 1 the same as Qwen3-30B-A3B-Thinking?
It is built on it. Nomos 1 is a post-trained version of Qwen3-30B-A3B-Thinking-2507 with the same architecture, and in the same harness it scored 87 of 120 on the Putnam against the base model's 24.
Does Nomos 1 write formal proofs in Lean?
No. Nomos 1 writes proofs in natural language, the way a contestant writes on paper. For proofs a computer can check in Lean 4, look at a formal prover such as Mistral's Leanstral 1.5.
Our verdict
Nomos 1 is worth running if you want an open model for competition maths and have the GPUs to host it.
Nous's post-training lifted a model that scored 24 of 120 on the Putnam to 87 in the same harness, and everything behind that result is free and public.
- Skip it for
- Anyone who wants a hosted chat or API, general coding help, or machine-checked Lean proofs.
- Check first
- Make sure you can serve about 122 GB of weights, or a community quantization, and let the harness run for hours per problem set.
MiroThinker
Hydrogen-1.0
Atria Dawn Preview
Hunyuan Vision 1.5 Thinking
Brave Ocelot
Biohub ESM