Stardust Flash-01
Stardust Flash-01
Our first open-weights model trained specifically for security research.
Stardust Flash-01 is an open-weights model post-trained on real vulnerabilities from our own private dataset. It is built as a worker model: a larger orchestrator hands it a focused security task, and it reads code, uses tools, tests ideas and confirms what it finds.
Why open weights#
Attackers already run open models locally, remove the safeguards and build on top. Defenders work under the tightest rules. Attacks that run at machine speed need defense at machine speed, so defenders need models they can run, control and improve inside the systems they protect.
At a glance#
| Base model | GLM-5.3-Flash |
| Method | Rank-256 LoRA across all experts and routers + full-parameter updates on 16 experts picked by activation, trained with GRPO |
| Training tasks | 150, built from 50 real vulnerability cases |
| Role | Worker model, run by an orchestrator |
| Held-out score | 66.7% (40/60) at $2.38 per full run |
Availability#
The open weights are on Hugging Face: Stardust-LLM/Stardust-Flash-01 (also the Try Local Model button on the home page). Flash-01 is not served through the Stardust API yet; it will appear in GET /v1/models when it is.