01 — A new kind of local

Intelligence that stays with you.

Luminas AI builds frontier-grade open models that run entirely on your own machine. Two generations, one idea: capability should not depend on someone else's data centre.

Current flagshipSora 3 · 9.4B
StatusPre-release
ModalitiesText + Vision
VRAM target6–8GB
First generationAstra 7
WeightsOpen · Hugging Face
02 — The upcoming flagship

Luminas Sora 3.

Sora 3 is the second generation of Luminas AI and a complete rethink of what a compact open model should be. 9.4 billion parameters, native vision, frontier-class reasoning and agentic tool use — engineered so the whole thing fits on a single everyday GPU.

Where the first generation answered questions, Sora 3 completes work: it reads a screenshot, plans an approach, edits files, runs the tests, inspects what came back and decides the next move.

Follow Sora 3 Full specifications
LUMINAS / SORA 3PRE-RELEASE
SORA3
PARAMETERS9.4B
MODALITYText + vision
RUNTIMELocal-first

What makes Sora 3 different

Six design decisions shape the model — each one chosen so capability per parameter, not parameter count, does the work.

01 — Architecture

Capability per parameter

9.4B is a deliberate ceiling. The model is tuned so quality comes from training and design rather than scale, keeping inference viable on consumer GPUs.

02 — Vision

Native multimodal input

Vision is part of the reasoning loop, not an add-on. Screenshots, interfaces, diagrams and design references are first-class context alongside text.

03 — Agency

Built to take action

Native function calling, code execution and tool use let the model plan, act, read real output and recover — across long, multi-step sessions.

04 — Footprint

6–8GB VRAM target

Designed around the hardware people already own, so the same model runs on a laptop GPU, a workstation or your own infrastructure.

05 — Openness

Open weights

Published openly through the Luminas-Ai organisation on Hugging Face, free to run, inspect, fine-tune and self-host.

06 — Privacy

Nothing leaves the machine

Local-first by default. Prompts, code and images stay on your hardware — no external inference call required.

03 — Capabilities

Small enough to run. Capable enough to build.

Modern software needs a model that understands context, decides what to do, uses tools, inspects results and keeps going.

01 — CREATE

Agentic coding

Plan multi-step development work, write code, inspect outputs and iterate across a whole project — with context intact.

Multi-fileTool useTests
02 — SEE

Vision in the loop

Interfaces, diagrams and screenshots become actionable context — visual input sits alongside text in the same session.

MultimodalUI readingDiagrams
03 — REASON

Frontier reasoning

Long-horizon inference and planning at the top of its weight class — in an open-weights package you can actually host.

PlanningLong contextRecovery
Input

Text + vision

Context from code, screenshots and instructions, in the same prompt.

Process

Reason →︎ act

Plan, call tools and inspect results inside one continuous loop.

Output

Useful work

Code, decisions and structured responses — not just paragraphs.

Runtime

Local-first

Designed around the hardware you already own.

$
→︎scanning repository…
→︎reading 14 relevant files
→︎planning changes
✓︎patch generated
✓︎tests executed · 42 passed
���reviewing diff
✓︎ready for next tool call
04 — Workflow

From prompt to workflow.

Sora 3 sits inside real software loops: understand the goal, choose an action, call a tool, read what happened, decide what comes next.

  • Plan before acting
  • Use tools in context
  • Recover and iterate from real outputs
  • Stay coherent across long, multi-step sessions
InputScreenshot →︎ structure

It reads the interface

ReasoningContext mapped

Then maps what matters

05 — Hardware & model profile

Frontier-class workflows. Local-first footprint.

Practical intelligence: a model you can experiment with and deploy without data-centre-class hardware.

9.4Billion parameters
6GB minimum VRAM
2Input modalities
100%Yours to run
Parameters
9.4B
Compact footprint, optimised for capability per parameter.
VRAM target
6–8GB
Runs on everyday GPUs. Memory varies by runtime and quantisation.
Modalities
Text + Vision
Visual context and language in a single model session.
Primary focus
Agentic coding
Multi-step reasoning, tool use and software workflows.
Reasoning
Frontier-class
Long-horizon planning and inference at the top of its weight class.
Weights
Open
Hosted openly on Hugging Face under the Luminas-Ai organisation.
Tool use
Native
Function calling, code execution and developer APIs out of the box.
Deployment
Local · Self-host
Run on your own machine or your own infrastructure.
Platforms
Win · macOS · Linux
Standard local runtimes on all three desktop platforms.
Availability
Pre-release
Launch announced through the Luminas AI Hugging Face page.
06 — Where Sora 3 stands

Pre-release, in the open.

No invented milestones and no placeholder numbers — here is exactly where the model is today.

Now

Pre-release

Architecture and capability set are locked: 9.4B, text + vision, agentic tool use, local-first.

In progress

Evaluation

The full benchmark suite is being run across coding, reasoning and vision tasks.

Next

Open weights

Weights published through the Luminas-Ai organisation on Hugging Face.

Then

Results published

Benchmarks land on this page the moment they are complete and verified.

07 — The first generation

Astra 7 started it. Sora 3 continues it.

Luminas Astra 7 was the first model Luminas AI ever shipped, and it set the principle everything since has followed: open weights, local execution, no dependency on hosted inference. It is still published on Hugging Face, still runnable, and still supported for the projects built around it.

First generation · Available

Luminas Astra 7

A language-focused open model built to prove a simple point — that a small, carefully trained model on ordinary hardware is enough for real work. Astra 7 remains the right choice if you have an existing integration, a tuned prompt stack or a deployment you do not want to disturb.

GenerationFirst
FocusLanguage
WeightsOpen
StatusAvailable · Maintained
Best forExisting projects
View Astra 7
Second generation · Upcoming

Luminas Sora 3

The flagship. Everything Astra 7 did, plus vision, frontier-class reasoning and native agentic tool use — in a 9.4B footprint still sized for a single consumer GPU. Start new work here.

GenerationSecond
Parameters9.4B
ModalitiesText + Vision
VRAM target6���8GB
StatusPre-release
Sora 3 · Coming Soon
Dimension
Astra 7
Sora 3
Generation
First — the original release
Second — current flagship
Input
Text
Text and vision in one session
Reasoning
General-purpose language reasoning
Frontier-class, long-horizon planning
Agentic work
Single-answer oriented
Native tool use, code execution, iteration
Deployment
Local, open weights
Local, open weights, 6–8GB VRAM target
Recommended for
Existing Astra-based projects
All new builds
08 — Evaluation

Benchmarks coming soon.

The scorecard stays empty until the evaluation suite is complete — no placeholder numbers, no invented rankings, just published results when they are ready.

Evaluation in progress

Coverage is planned around the areas Sora 3 is built for: agentic coding, tool use, reasoning, vision, code generation and long-context tasks.

Agentic codingTool useReasoning VisionCode generationLong-context
09 — The details

Questions, answered.

What is Luminas AI?
Luminas AI is a research and product organisation building frontier-grade open language models that run locally on everyday hardware. The family includes Luminas Astra 7, the first generation, and the upcoming Luminas Sora 3.
What is Luminas Sora 3?
Sora 3 is the upcoming 9.4B multimodal flagship. It combines vision, frontier-level reasoning and agentic coding, and is designed to run locally on GPUs in the 6–8GB VRAM range.
What can Sora 3 actually do?
Plan and carry out multi-step development work, call tools and execute code, read screenshots and interfaces as context, reason over long horizons, and recover when a step fails — all inside one continuous local session.
Can Sora 3 run on a 6GB or 8GB GPU?
Yes. Sora 3 is designed around the 6–8GB VRAM range. Actual memory use depends on runtime, quantisation, context length and configuration.
What is Luminas Astra 7?
Astra 7 is the first model Luminas AI released — a language-focused open-weights model built for local execution. It established the local-first, open-weights approach and is still published on Hugging Face for projects built around the first generation.
Should I move from Astra 7 to Sora 3?
If you have a working Astra 7 deployment, there is no pressure to move — it remains available and maintained. For new work, start on Sora 3: it adds vision, substantially stronger reasoning and native agentic tool use in the same local-first footprint.
Where can I download the models?
Both Sora 3 and Astra 7 are published through the Luminas-Ai organisation on Hugging Face, with open weights you can run, inspect and self-host.
When is Sora 3 releasing?
Sora 3 is in pre-release while evaluation is completed. The launch will be announced on the Luminas AI Hugging Face page.

Make space for better thinking.

9.4B parameters. Vision. Frontier reasoning. Agentic coding. Designed for the hardware you already have.

Sora 3 · Coming Soon Visit Luminas AI
Luminas Sora 3 · Pre-release · Open weights