dusersad12/VertexNova-PreviewRepo

VertexNova

---

1. Introduction

VertexNova is the newest snapshot in our open-weight language model series, post-trained with a longer reasoning curriculum and a denser preference-optimization schedule than everything we shipped before. Across math, code and general knowledge probes it now sits comfortably ahead of our earlier releases, and it closes most of the remaining gap to several frontier-scale models while staying small enough to serve on a single accelerator.

The headline change is reasoning depth. On an internal AIME-style probe the pass rate moved from 61.9% in the previous snapshot to 84.6%, and average token usage per question grew from roughly 9K to 21K, which tells us the model actually spends its extra compute thinking. We also recorded fewer unsupported factual claims and noticeably steadier tool use in agent-style evaluations.

2. Evaluation Results

Comprehensive Benchmark Results

	Benchmark	ModelA	ModelB	ModelA-v2	VertexNova
Core Reasoning Tasks	Math Reasoning	0.521	0.548	0.664	0.691
	Logical Reasoning	0.589	0.612	0.726	0.750
	Common Sense	0.604	0.655	0.713	0.742
Language Understanding	Reading Comprehension	0.556	0.584	0.690	0.713
	Question Answering	0.518	0.546	0.660	0.683
	Text Classification	0.623	0.651	0.742	0.768
	Sentiment Analysis	0.572	0.599	0.674	0.690
Generation Tasks	Code Generation	0.537	0.564	0.668	0.691
	Creative Writing	0.566	0.549	0.673	0.698
	Dialogue Generation	0.574	0.608	0.704	0.729
	Summarization	0.593	0.629	0.694	0.717
Specialized Capabilities	Translation	0.602	0.641	0.703	0.728
	Knowledge Retrieval	0.527	0.559	0.661	0.686
	Instruction Following	0.611	0.637	0.714	0.739
	Safety Evaluation	0.614	0.592	0.706	0.730

Overall Performance Summary

VertexNova posts its largest gains on the reasoning-adjacent categories, and it stays competitive on every generation-oriented task in the suite.

3. Chat Website & API Platform

An interactive demo plus hosted endpoints for VertexNova are available through our project page; bring your own API key and the usual chat completions interface works out of the box.

4. How to Run Locally

VertexNova loads through the standard transformers AutoModel/AutoTokenizer interface — no custom code is required. We recommend greedy decoding with temperature 0.6 for math-heavy workloads, and a short system prompt that states the current date for time-sensitive queries.

5. License

The VertexNova weights and the accompanying evaluation code are released under the Apache 2.0 license.

6. Contact

Questions and bug reports are welcome on our issue tracker, or reach the team at contact@vertexnova.ai.