richardyoung/ nomos-1:iq4_xs

14 7 months ago

A 30B-parameter (Mixture-of-Experts) model specialized for **mathematical problem-solving and proof-writing in natural language**, packaged for Ollama with a 256K context window. 

ollama run richardyoung/nomos-1:iq4_xs

Details

7 months ago

517953b90bc9 · 17GB ·

model
qwen3moe
·
30.5B
·
IQ4_XS
params
{ "stop": [ "<|im_start|>", "<|im_end|>" ] }
template
{{ if .System }}<|im_start|>system {{ .System }}<|im_end|> {{ end }}{{ if .Prompt }}<|im_start|>user

Readme

Nomos-1

A 30B-parameter (Mixture-of-Experts) model specialized for mathematical problem-solving and proof-writing in natural language, packaged for Ollama with a 256K context window.

🚀 Overview

nomos-1 is an Ollama distribution of Nomos-1, a model released by NousResearch that specializes the Qwen/Qwen3-30B-A3B-Thinking-2507 foundation model for deep mathematical reasoning, problem-solving, and proof-writing in natural language. According to the upstream model card, Nomos-1 was trained in collaboration with Hillclimb AI and is designed to be used alongside the Nomos Reasoning Harness, which was open-sourced concurrently.

This packaging exposes the model as ready-to-run GGUF quantizations on Ollama. The underlying architecture is Qwen3 MoE (qwen3moe) with 30.5B total parameters and an advertised 256K context window. It uses the ChatML prompt format (<|im_start|> / <|im_end|>).

  • Base lineage: Qwen/Qwen3-30B-A3B-Thinking-2507 (NousResearch/nomos-1)
  • Architecture: Qwen3 MoE (qwen3moe)
  • Total parameters: 30.5B
  • Context window: 256K (as advertised on the model’s Ollama page)
  • License: Apache 2.0 (per the upstream NousResearch model card)

Note: Per the upstream model card, it is recommended to run Nomos-1 without a system prompt.

📊 Benchmarks

The upstream model card reports the following result on Putnam 2025, measured when the model is wrapped in the Nomos Reasoning Harness:

Model Putnam 2025 (with Nomos Reasoning Harness)
Nomos-1 87 / 120
Qwen3-30B-A3B-Thinking-2507 (base) 24 / 120

These figures are from the official NousResearch model card and reflect performance with the reasoning harness; results without the harness are not stated here.

🎯 Key Features

  • 🧮 Math-specialized, tuned for problem-solving and rigorous proof-writing in natural language.
  • 🧠 Reasoning (“Thinking”) base, built on Qwen3-30B-A3B-Thinking-2507, a chain-of-thought–oriented MoE model.
  • 📏 256K context, long-context window suitable for lengthy problems, multi-step derivations, and large supporting material.
  • ⚡ Efficient MoE, ~30.5B total parameters with sparse activation (A3B), giving strong reasoning at a fraction of the active compute of a dense 30B model.
  • 🧩 Harness-ready, designed to pair with the open-source Nomos Reasoning Harness.

🏷️ Available Versions

Tag Quantization Size Context
nomos-1:Q4_K_M Q4_K_M 19 GB 256K
nomos-1:iq4_xs IQ4_XS 17 GB 256K

Q4_K_M generally offers higher fidelity; iq4_xs is slightly smaller and lighter on memory. Sizes are as listed on the model’s Ollama page.

💻 Quick Start

# Pull and run (Q4_K_M)
ollama run richardyoung/nomos-1:Q4_K_M

# Or the smaller IQ4_XS build
ollama run richardyoung/nomos-1:iq4_xs

# Example: ask a math problem (no system prompt, per upstream guidance)
ollama run richardyoung/nomos-1:Q4_K_M "Prove that the sum of the first n odd positive integers equals n^2."

For best results, follow the upstream recommendation and run without a system prompt, and consider pairing the model with the Nomos Reasoning Harness for the strongest results.

🛠️ Use Cases

  • 📐 Mathematical problem-solving (competition-style and general).
  • ✍️ Writing and checking natural-language proofs.
  • 🎓 Tutoring and step-by-step explanation of mathematical reasoning.
  • 🔬 Research and experimentation with reasoning harnesses on long-context math tasks.

📋 System Requirements

These 4-bit builds are ~17–19 GB on disk. Practical guidance:

Setup Notes
Recommended A GPU with 24 GB+ VRAM (e.g. RTX 3090⁄4090) for comfortable inference at moderate context.
Minimum ~20 GB combined VRAM/RAM to load the weights; expect slower CPU/offloaded inference below that.
Long context Using the full 256K window substantially increases memory use for the KV cache, additional headroom is required.

Actual requirements scale with the context length you use and your quantization choice.

🔧 Technical Details

Property Value
Base model Qwen/Qwen3-30B-A3B-Thinking-2507
Upstream release NousResearch/nomos-1
Architecture Qwen3 MoE (qwen3moe)
Total parameters 30.5B
Context window 256K (advertised)
Prompt format ChatML (`<
Quantizations Q4_K_M (19 GB), IQ4_XS (17 GB)
License Apache 2.0 (per upstream card)
Specialization Mathematical problem-solving & proof-writing

⚠️ Disclaimer

This is a community redistribution of NousResearch’s Nomos-1 in GGUF/Ollama form. It is provided “as is”, without warranty of any kind. Nomos-1 is specialized for mathematical reasoning; like all language models it can produce incorrect or incomplete reasoning, and its outputs, including proofs, should be independently verified before being relied upon. Use is subject to the upstream Apache 2.0 license and the licenses/terms of the underlying Qwen base model. The benchmark figures shown are reproduced from the upstream model card and were obtained with the Nomos Reasoning Harness; your results may differ.

🙏 Acknowledgments

  • NousResearch, for releasing Nomos-1.
  • Hillclimb AI, collaborator on training, per the upstream model card.
  • Qwen Team (Alibaba), for the Qwen3-30B-A3B-Thinking-2507 base model.
  • Ollama and llama.cpp, for the runtime and GGUF tooling that make local inference possible.

Built & maintained by Richard Young · DeepNeuro