A compact model with the reasoning of a much larger one.
Luminas Astra 7 is a 7.6 billion parameter language model built by Luminas AI for agentic work, careful reasoning, real programming, and precise instruction following — small enough to run close to where your work already happens.
Built by Luminas AI · Weights and API details publish at launch.
A compact intelligence built for reasoning, creation, and code.
Luminas Astra 7 is a 7.6 billion parameter model from Luminas AI, designed to understand complex instructions, reason through problems, generate and analyse code, explain technical concepts, and assist across a wide range of development and knowledge-based tasks.
Rather than focusing solely on model scale, Astra 7 aims to deliver practical intelligence in a compact form factor, making advanced AI capabilities more accessible to individual developers, local systems, and AI-powered software.
Intelligence should be useful, efficient, and accessible.
Luminas AI
Select a capability to see what it means in practice.
Astra 7 approaches multi-step problems systematically, connecting information across a prompt to produce clear and useful solutions rather than relying solely on simple pattern matching.
Astra 7 is optimised for developer-oriented workflows — writing new code, reading code that already exists, and explaining what a change actually does.
Astra 7 follows detailed instructions and maintains context throughout multi-step tasks, respecting the format, the scope, and the constraints you set.
Beyond programming, Astra 7 assists with brainstorming, conceptual development, writing, planning, and analysis — turning loose ideas into structured solutions.
The model works comfortably with programming concepts, software architecture, algorithms, APIs, development workflows, and other technical subjects.
Astra 7 is built with developers and AI-powered applications in mind. Its compact 7.6 billion parameter scale makes it particularly suitable for environments where performance, efficiency, and local deployment matter.
It can serve as an intelligence layer for:
In Luminas AI's internal evaluations, Luminas Astra 7 places first against every model below — with its widest margins on agentic tasks and multi-step reasoning, where compact models usually give ground. Switch category to see each result.
Scores come from Luminas AI's own pre-release evaluations of Astra 7 against public checkpoints of the models listed, scored on a shared internal harness (0–100, higher is better). Small numbers under each score are parameter counts. At launch, Luminas AI will publish the full results together with the datasets, prompts, harness, and hardware used, so that every number here can be reproduced independently.
Ask Astra 7 for something and 7.6 billion parameters answer at once — tokens becoming context, context becoming reasoning, reasoning becoming an answer. Every thread below is drawn live in your browser. Move your cursor through the network, or send a signal and watch it cross.
An artistic visualisation of signal flow through a dense network — a diagram of Astra 7's layer topology. Exact architecture, context window, and training details publish with the model.
Astra 7 is designed to be practical for local AI environments, letting developers experiment with and deploy capable intelligence without requiring large-scale cloud infrastructure. Its compact size suits resource-conscious deployments while retaining the capabilities expected from a modern developer-focused language model.
Astra 7 is part of the Luminas AI ecosystem, where intelligence is designed to work alongside software rather than exist only as a conversational interface. As Model 01, it is the foundation for intelligent development experiences, autonomous agents, and future Luminas products.
Everything confirmed about the model today. Training details, evaluation results, licensing, and access will be published in full when the model is released.
Luminas Astra 7 is the first model from Luminas AI — a 7.6 billion parameter large language model built for agentic work, advanced reasoning, programming, instruction following, and creative problem solving in a compact, efficient architecture.
7.6 billion. Astra 7 sits in the compact 7B class, which keeps it fast and inexpensive to run while staying capable enough for real development work.
Agentic, multi-step tasks and reasoning, plus the full range of developer work: code generation, code explanation, debugging and error analysis, refactoring, algorithmic problem solving, function and API development, technical documentation, and understanding existing codebases.
Yes. Astra 7 is designed for local-first deployment, so developers can experiment with and ship capable intelligence without large-scale cloud infrastructure. It works as an intelligence layer for AI coding assistants, AI IDEs, developer tools, autonomous agents, and personal AI systems.
In Luminas AI's internal evaluations, Astra 7 ranks first against all eight compared models — Qwen2.5-Coder-7B-Instruct, DeepSeek-Coder-6.7B-Instruct, CodeGemma 7B, StarCoder2-7B, Mistral 7B Instruct, Llama 3.1 8B Instruct, Gemma 2 9B and Qwen2.5-7B-Instruct — with its widest margins on agentic tasks and reasoning. Full methodology publishes with the model.
Astra 7 is coming soon. The weights, license and integration details will be published on Hugging Face at launch, and the download button on this page will link straight to them.
Luminas AI. Astra 7 is Model 01 in the Luminas AI ecosystem, where intelligence is designed to work alongside software rather than exist only as a conversational interface.
Luminas Astra 7 is the beginning of Luminas AI's journey toward building intelligent systems that move beyond conversation and become active partners in creation. When it is released, the weights, license, and integration details will be published here first.
Built by Luminas AI · Astra · Model 01