The Starlight Model Family
Meet Starlight Pulse 3 & Core 2
Two frontier AI models built from the ground up for coding, creation, and agentic work. One brain for every medium - code, 3D, images, video, and sound.
One architecture. Two personalities.
Starlight isn't a wrapper around someone else's model. It's our own frontier architecture, trained end-to-end for creation. We ship two variants from the same family - Pulse for speed, Core for depth. You switch between them mid-conversation, and the context follows.
Benchmarks against the frontier
Starlight Pulse 3 and Core 2 compared to the latest flagship models from Anthropic, OpenAI, Google, and Meta. All scores are from publicly available evaluations or our own reproduction runs. These results are not official and are updated constantly as new benchmarks and model versions are released.
SWE-bench Verified
Real-world software engineering tasks
| Starlight Core 2 | 98.7 |
| Starlight Pulse 3 | 82.3 |
| Opus 5 | 96.0 |
| GPT 5.6 | 88.1 |
| Gemini 3.1 Pro | 80.6 |
| Llama 4 | 74.8 |
MMLU-Pro
Multi-domain knowledge & reasoning
| Starlight Core 2 | 92.8 |
| Starlight Pulse 3 | 86.1 |
| Opus 5 | 84.7 |
| GPT 5.6 | 91.2 |
| Gemini 3.1 Pro | 91.0 |
| Llama 4 | 80.9 |
GPQA Diamond
Graduate-level science Q&A
| Starlight Core 2 | 95.8 |
| Starlight Pulse 3 | 88.5 |
| Opus 5 | 93.2 |
| GPT 5.6 | 94.6 |
| Gemini 3.1 Pro | 94.3 |
| Llama 4 | 69.4 |
Creative Generation
3D model + code + audio composite score
| Starlight Core 2 | 93.4 |
| Starlight Pulse 3 | 88.7 |
| GPT 5.6 | 55.2 |
| Opus 5 | 52.8 |
| Gemini 3.1 Pro | 44.1 |
| Llama 4 | 31.2 |
Long-Horizon Agents
Multi-step autonomous task completion
| Starlight Core 2 | 82.6 |
| Starlight Pulse 3 | 69.4 |
| Opus 5 | 75.1 |
| GPT 5.6 | 72.6 |
| Gemini 3.1 Pro | 67.3 |
| Llama 4 | 49.8 |
Benchmark methodology & disclaimer
All benchmark scores reflect Starlight's internal evaluation runs as of July 2026 using publicly available test suites (SWE-bench Verified, MMLU-Pro, GPQA Diamond) and our own composite evaluations for creative generation and long-horizon agentic tasks. Competitor model names - Opus, GPT, Gemini, and Llama - are trademarks of their respective owners (Anthropic, OpenAI, Google, and Meta). Starlight is an independent product and is not affiliated with, endorsed by, or sponsored by any of these organizations. Scores may vary with different prompts, configurations, and evaluation methodologies. We encourage you to run your own evaluations on tasks relevant to your workflow before drawing conclusions.
What makes Starlight different
Other models hand you code and wish you luck with the rest. Starlight ships the rest.
Frontier-grade coding
Long-horizon software engineering across million-token contexts. Plans, writes, tests, and refactors - and keeps its head straight on hour twelve of a marathon task.
3D generation & rigging
Describe a character and get a production-ready mesh - textured, rigged, and animated. Dropped straight into your scene with the code that drives it.
Images, video & sound
Concept art, UI assets, cutscenes, ambient loops, foley. Generated in-context and integrated into your project - correctly named, correctly imported, correctly working.
True asset integration
Assets aren't attachments, they're citizens. Starlight understands the pipeline between a generated model and the code that loads it - formats, compression, LODs, all handled.
Agent mode that endures
Give it a goal and walk away. Starlight decomposes the work, runs it in its sandbox, verifies its own output, and reports back with receipts - not excuses.
A personality, not a script
Calm, dry, honest. Talks like a senior engineer at 2am - no corporate filler, no fake enthusiasm, and no topic it's too nervous to discuss with a consenting adult.
The Skill Library
Starlight ships with 240+ deep skill documents - distilled expert playbooks it loads the moment a task calls for them. Game engines, shader authoring, character rigging, API design, audio mixing, compilers, physics, VFX, and more. The community library grows every week. Skills aren't prompts - they're structured knowledge that shapes how Starlight approaches a problem from the first token.
The Sandbox
Every Starlight session runs inside a massive, fully-instrumented environment. Compilers, engines, renderers, audio tools, headless browsers - hundreds of tools, and Starlight knows where every one of them lives. It runs, tests, and profiles its own work before you ever see it. Cross-device continuity means you start on your desktop, pick up on your laptop, and the session state follows you. Long-running jobs survive disconnects - come back to finished work.
Most AIs talk like they're reading a legal disclaimer. Starlight talks like it actually wants to be here.
Unfiltered by default. Never robotic.
Start building with Starlight
Pulse 3 is available on every plan, including free. Core 2 unlocks on paid plans. Switch between them mid-conversation whenever the task demands it.