Nexora-Vector-v0.1: A New Direction in AI-Assisted Vector Graphics

The idea of describing an image in plain English and receiving a usable vector graphic in return is something designers and developers have long wanted. Most AI image generation tools produce raster outputs, which are difficult to scale, edit, or integrate into vector-based workflows. Vector graphics, by contrast, are infinitely scalable, lightweight, and directly usable in web and print design.

Nexora-Vector-v0.1 is an attempt to solve this problem. Released by ArkAiLabs, a sub-brand of ArkDevLabs, this model takes natural language instructions and outputs structured SVG code. It is an early-stage release, clearly labeled as beta, but it represents a meaningful step toward making text-to-vector generation a practical reality.

The concept and development of this model originated with JackMa, co-founder of ArkDevLabs, who envisioned a lightweight AI system capable of bridging natural language and structured vector output in a way that is accessible for both research and rapid prototyping.

What Is Nexora-Vector-v0.1?

Nexora-Vector-v0.1 is a supervised fine-tuned language model based on Qwen3-4B. Rather than generating prose or answering questions, this model is adapted specifically to produce SVG markup from textual prompts.

SVG, or Scalable Vector Graphics, is an XML-based format for describing two-dimensional graphics. Because SVG is code rather than pixels, it can be edited, scaled, and integrated into web and design workflows directly. A model that can reliably generate valid SVG from a prompt would have significant practical value in design tooling, rapid prototyping, and educational applications.

Nexora-Vector-v0.1 is the first model in the Nexora Vector series, and is intended as a research and experimentation release rather than a production-grade tool.

The Idea Behind the Model

The vision for Nexora-Vector came from JackMa, who wanted to explore what a small, focused language model could do when fine-tuned specifically for structured creative output.

Rather than building a general-purpose model, the goal was to create something purpose-built: a model that understands prompts describing simple visuals and translates them directly into renderable SVG code. This reflects a broader philosophy at ArkAiLabs of building lightweight, practical, and focused AI systems rather than attempting to compete with large general-purpose models.

The Nexora Vector series is part of the Nexora initiative under ArkAiLabs, which explores creative and structured AI output generation as a distinct research direction.

Architecture and Training

The model is built on Qwen3-4B, a capable open-weight base model from the Qwen team. Fine-tuning was performed using supervised learning on curated prompt-SVG pairs.

Parameter	Details
Fine-tuning Method	Supervised Fine-Tuning (SFT)
Dataset Composition	Curated prompt-SVG pairs
Dataset Size	~1,500 samples
Training Objective	Structured output generation for SVG formats

The training dataset is deliberately scoped. Around 1,500 prompt-SVG samples were used for this initial release. This is a small dataset by modern standards, and the team acknowledges that this contributes to some of the model's current limitations. Expanding the dataset is a core part of the roadmap for future versions.

What the Model Can Do

Nexora-Vector-v0.1 is designed to handle straightforward vector generation tasks. Its strengths include:

• Generating SVG markup for simple geometric shapes and compositions
• Producing lightweight icons and minimal design assets
• Assisting in early-stage prototyping for vector-based design workflows
• Supporting research into structured output generation from language models

The model performs best when prompts are concise and focused on simple visual elements. A prompt like "a blue circle centered on a white background" is more likely to yield a usable output than a prompt describing a complex multi-element scene.

Current Limitations

This is a beta release, and the team at ArkAiLabs has been transparent about where the model falls short at this stage.

The most significant issue is a high hallucination rate. The model can produce SVG outputs that are malformed, non-renderable, or structurally incorrect. This is expected behavior for a model trained on a limited dataset, and it means that all outputs need to be validated before use.

Complex scenes are also a weak point. Prompts that describe multiple interacting elements, fine spatial relationships, or detailed compositions tend to produce inconsistent results. The model's generalization across diverse prompts is limited by the size and scope of the training data.

For these reasons, Nexora-Vector-v0.1 is not suitable for automated pipelines or production-grade design workflows where SVG correctness is required without human review.

How to Use It Effectively

Getting useful results from Nexora-Vector-v0.1 requires some awareness of its strengths and limitations. A few practical recommendations:

Keep prompts simple and specific. The more focused and concrete the prompt, the more likely the model is to produce a valid output. Avoid vague descriptions or multi-scene compositions in a single prompt.

Validate all outputs. Treat every output as a draft. SVG syntax checking and manual review should be part of any workflow that uses this model.

Use iterative prompting. If an initial prompt produces a poor result, refining and adjusting the prompt across multiple attempts often yields better outputs.

Expect imperfections. This is a first release. The goal at this stage is research and exploration, not production-ready output.

Quantized Versions

ArkAiLabs has made quantized versions of Nexora-Vector-v0.1 available through Open4bits, its dedicated quantization project, to support efficient local inference across different hardware platforms.

Two official quantized formats are available:

GGUF — suitable for local inference on Windows, Linux, and macOS using tools such as llama.cpp, Ollama, or LM Studio. Multiple quantization levels are available including Q2_K, Q4_K_M, Q6_K, and Q8_0.

Download: Open4bits/nexora-vector-v0.1-GGUF

MLX 4-Bit — optimized for Apple Silicon (M1/M2/M3/M4) via the MLX framework. This version is the recommended choice for users running inference on Apple hardware.

Download: Open4bits/nexora-vector-v0.1-mlx-4Bit

Both releases are maintained by Open4bits under ArkAiLabs and are designed to make the model accessible for local use without requiring high-end hardware.

Evaluation and Roadmap

Nexora-Vector-v0.1 has not yet undergone formal benchmark evaluation. The current assessment of the model is qualitative, based on manual testing of SVG generation tasks. Formal evaluation metrics are planned for future releases and will cover SVG validity rate, structural correctness, prompt adherence, and visual consistency across similar inputs.

The roadmap for the Nexora Vector series includes several significant improvements:

• An expanded and more diverse training dataset
• Improved SVG syntax correctness and validity rates
• Reduced hallucination rates
• Better natural language understanding for complex or multi-element prompts
• Support for richer vector compositions
• A formal benchmark evaluation suite

The foundation laid by this initial release will inform the direction of these improvements.

Licensing and Community

Nexora-Vector-v0.1 is released under the Apache License 2.0, which permits use, modification, and distribution of the model in accordance with the license terms.

ArkDevLabs maintains an active Discord community where users can follow updates, share results, and engage in discussion around the Nexora project. You can join at discord.gg/mwdrgYbzuG.

The model is built on Qwen3-4B, developed by the Qwen team, and the ArkAiLabs team acknowledges the broader open-source AI community whose work makes projects like this possible.

Conclusion

Nexora-Vector-v0.1 is an honest and focused first release. It does not try to be a general-purpose model or overstate what it can deliver. It is a specialized tool for a specific and genuinely useful task, built by a small team with a clear vision for where the technology needs to go.

The idea, developed by JackMa and brought to release by ArkAiLabs, reflects a practical approach to AI development: start with a defined problem, build something lightweight and purposeful, release it openly, and improve it iteratively based on real-world use.

For researchers, developers, and designers interested in the frontier of text-to-vector generation, Nexora-Vector-v0.1 is worth exploring. Just keep a validator nearby.