Guides / Math
Best small open models for math
Updated November 2026
As of November 2026, the best option is Kolibri-1 by Aleph Alpha, with an AIME score of 96.0, followed by Ternary Bonsai 2 27B (95.8) and Nanbeige4.3 33B (95.2).
Ranked by AIME (American Invitational Mathematics Examination), reported or estimated. The exam year differs between models; see each model page.
| # | Model | Lab | AIME score | Index | Params | Memory (4-bit) | API / 1M | License |
|---|---|---|---|---|---|---|---|---|
| 1 | Kolibri-1 | Aleph Alpha | 96.0 | 68.3 | 78.1B (3.5B active) | ~49.6 GB | self-host | Apache-2.0 |
| 2 | Ternary Bonsai 2 27B | PrismML | 95.8 | 74.6 | 27.4B | ~18.2 GB | self-host | Apache-2.0 |
| 3 | Nanbeige4.3 33B | Nanbeige (Kanzhun) | 95.2 | 77.8 | 33.1B | ~23 GB | self-host | Apache-2.0 |
| 4 | Muse Glimmer 30B | Meta | 94.7 | 72.6 | 29.6B | ~19.2 GB | $0.53 | Apache-2.0 |
| 5 | Qwen3.6 27B | Alibaba (Qwen) | 94.1 | 74.0 | 27B | ~17.9 GB | $0.90 | Apache-2.0 |
| 6 | Qwen3.6 35B-A3B | Alibaba (Qwen) | 92.7 | 72.2 | 35B (3B active) | ~22.6 GB | $0.031 | Apache-2.0 |
| 7 | Qwen3.5 27B | Alibaba (Qwen) | 92.6 | 70.6 | 27B | ~17.9 GB | $0.69 | Apache-2.0 |
| 8 | gpt-oss-20b | OpenAI | 91.7 | 64.0 | 20.9B (3.6B active) | ~13.8 GB | $0.020 | Apache-2.0 |
| 9 | GLM-4.7-Flash | Z.ai (GLM) | 91.6 | 66.1 | 31B (3B active) | ~20.4 GB | $0.15 | MIT |
| 10 | Spark-X2.6 4B | SparkLLM (XHToken) | 91.2 | 67.5 | 4.1B | ~3.4 GB | self-host | Apache-2.0 |
| 11 | Qwen3.5 35B-A3B | Alibaba (Qwen) | 91.0 | 69.9 | 35B (3B active) | ~22.6 GB | $0.24 | Apache-2.0 |
| 12 | Spark-X2.5 4B | SparkLLM (XHToken) | 90.7 | 65.5 | 4.1B | ~3.4 GB | self-host | Apache-2.0 |
| 13 | Gemma 4 31B | 89.2 | 70.1 | 30.7B | ~21.9 GB | $0.16 | Apache-2.0 | |
| 14 | Granite 4.2 30B | IBM (Granite) | 89.2 | 67.0 | 30B | ~21.4 GB | self-host | Apache-2.0 |
| 15 | Nemotron 3 Nano 30B-A3B | NVIDIA | 89.1 | 62.6 | 31.6B (3.5B active) | ~20.3 GB | $0.10 | NVIDIA Open Model |
Memory covers the weights only; long contexts need more. Scores are mostly lab-reported, so gaps of 1–2 points are noise. See the methodology, or filter the full leaderboard yourself.
Head to head
Questions
What is the best small open-weight model for math?
As of November 2026, the best option is Kolibri-1 by Aleph Alpha, with an AIME score of 96.0, followed by Ternary Bonsai 2 27B (95.8) and Nanbeige4.3 33B (95.2).
How is this list ranked?
Ranked by AIME (American Invitational Mathematics Examination), reported or estimated. The exam year differs between models; see each model page.
Which of the top five needs the least memory?
Qwen3.6 27B, at about 17.9 GB with 4-bit quantization.
Which is the cheapest to use through an API?
gpt-oss-20b, at $0.020 per million tokens from the cheapest provider on OpenRouter.
Other guides
- Best small open models for coding
- Best LLMs that run on an 8 GB GPU or laptop
- Best LLMs that run on a 16 GB GPU
- Best LLMs for a 24 GB GPU (RTX 4090, RTX 3090)
- Best tiny LLMs for phones and edge devices
- Best cheap LLM APIs (open models)
- Best small open vision-language models
- Best open models with an Apache 2.0 or MIT license
- Best European open language models
- Best American open models under 100B
- Best fully open models (open data and code)