Announcements
Nova: Introducing Our Latest Suite of LLMs
At Nova, our mission is to redefine AI capabilities. We are pleased to introduce our latest range of large language models (LLMs), which deliver a balance of speed, quality, and specialised capabilities tailored to a variety of needs.
Try Nova
Our new models bring exceptional proficiency in reasoning and mathematics, making them a powerful addition to any AI toolkit.
Introducing Nova-Instant, Nova-Air, and Nova-Pro
Nova's new LLMs — Nova-Instant, Nova-Air, and Nova-Pro — are developed to cater to different requirements, ranging from fast, cost-effective solutions to high-performance models. Each model within the Nova suite is optimised for distinct purposes, ensuring a suitable tool is available for every task.
Benchmark Performance
The Nova models demonstrate strong capabilities across several key benchmarks, showcasing improvements in general, reasoning, mathematics, coding tasks. The following are some notable results:
- MMLU Benchmark: Nova-Air achieved 86.1% and Nova-Pro achieved 88.8% on five-shot MMLU benchmarks, highlighting their robust general reasoning capabilities.
- Reasoning: Nova-Pro excelled with a 97.2% outcome on ARC grade-school science questions.
- Mathematics - Basic Arithmetic (GSM8K): Nova-Pro reached 96.9% in five-shot, chain-of-thought settings.
- Coding - HumanEval: Nova-Pro achieved 91.8%, showing strong human-like code generation capabilities.
Notably, Nova-Pro led the field with an impressive 88.8% on the MMLU benchmark and excelled in both coding (91.8% on HumanEval) and scientific reasoning (97% on ARC-C). Nova-Instant and Nova-Air also displayed strong performances, proving to be highly effective in a range of applications.
In comparison with other leading models, Nova-Pro consistently outperforms across various metrics, particularly excelling in reasoning and mathematics. This positions Nova-Pro as a superior model compared to models like GPT-4 Omni and Claude 3.5 Sonnet.
In conclusion, the Nova suite maintains a strong balance between performance, speed, and versatility, making it well-suited for a wide array of tasks while delivering industry-leading performance in essential benchmarks.
Benchmarks
A comparative analysis of the performance of popular State-of-the-Art (SOTA) and Nova models across a diverse range of benchmarks.
| Capability | Benchmark | Nova-Instant | Nova-Focus (beta) | Nova Air | Nova Pro | GPT-4 OMNI | Claude 3.5 Sonnet |
|---|---|---|---|---|---|---|---|
| General | MMLU-Pro | 50.0% | 64.6% | 69.8% | 74.2% | 72.5% | 76.1% |
| General | MMLU | 73.5% | 80.0% | 86.1% | 88.8% | 88.7% | 88.3% |
| Reasoning | ARC-C | 83.9% | 87.2% | 95.3% | 97.0% | 96.7% | N/A |
| Reasoning | GPQA | 33.6% | 45.5% | 48.0% | 52.8% | 53.6% | 59.4% |
| Math | GSM8K | 85.3% | 95.1% | 95.3% | 96.9% | 96.1% | 96.4% |
| Math | MATH | 52.1% | 80.0% | 70.5% | 74.2% | 76.6% | 71.1% |
| Code | HumanEval | 72.9% | 83.5% | 82.0% | 91.8% | 90.2% | 92.0% |
| Code | IFeval | 80.6% | 81.0% | 87.5% | 88.9% | 85.6% | N/A |
| Multilingual | MGSM | 70.5% | 82.7% | 87.2% | 91.9% | 90.5% | 91.6% |
What's Next?
Our objective is to develop low-cost, high-performance, reliable, and privacy-preserving Large Language Models (LLMs). In pursuit of these aims and the expansion of our Nova model series, we are currently preparing the release of a stable version of Nova-Focus. To enhance quality, we are experimenting with the Chain-of-Thought (CoT) reasoning approach for an upcoming model release, which has demonstrated impressive improvements in reasoning, mathematics, and coding capabilities. These advancements will position the model to compete with OpenAI's o1 models in benchmarks such as MMLU, MATH, and HumanEval, with additional benchmarks to follow.
Introducing LumiNova: State-of-the-art Image Generation
We are excited to introduce LumiNova, our advanced image generation model. LumiNova is based on a fine-tuned version of the Flux model, courtesy of Fal.ai, and leverages the latest advancements in image synthesis.
LumiNova expands the boundaries of visual creativity and is engineered to deliver high fidelity, diverse, and visually stunning outputs. Whether it is adhering closely to prompts or handling a wide variety of image styles and complexities, LumiNova offers unparalleled performance among contemporary models.
Benchmark Performance
LumiNova showcases exceptional performance across multiple categories, as outlined by FLUX.1 benchmarks:
- Visual Quality: LumiNova exhibits top-of-the-line visual quality with exceptional detail and refinement.
- Prompt Adherence: LumiNova consistently follows complex prompts with precision, providing consistent and reliable image outputs.
- Output Diversity: LumiNova offers a wide range of styles and outputs, supporting diverse creative needs.
- Size & Aspect Ratio Variability: LumiNova handles a range of aspect ratios and resolutions, offering flexibility in image size from 0.1 to 2.0 megapixels.
LumiNova brings fine-tuned excellence to the forefront of image generation, offering a comprehensive solution for high-quality, prompt-adherent, and diverse visual outputs powered by state-of-the-art technology.
Where to try Nova?
You can access Nova models at rubiks.ai/nova.