이미지: METAL LAB 생성
Summary
- Dev-automation startup Factory unveiled a locally run setup built on NVIDIA DGX Spark and the Nemotron 3.5 Lightning model
- The company says security-conscious organizations can now run their own "software factory" without moving source code or execution environments off-premises
- Nemotron 3.5 Lightning, an open model that already received day-one support from vLLM and others, runs across a wide range of hardware from DGX Spark up to large GPUs
- 발표 주체
- Factory (개발툴·에이전트 스타트업)
- 발표일
- 2026-08-11
- 구동 하드웨어
- NVIDIA DGX Spark
- 사용 모델
- NVIDIA Nemotron 3.5 Lightning
- 모델 구조
- 하이브리드 MoE, 전체 30B·활성 3B (배경 확인)
- 모델 컨텍스트 길이
- 최대 100만 토큰 (배경 확인)
- 핵심 가치 제안
- 소스코드·컨텍스트·실행을 조직 환경 밖으로 옮기지 않고 운영
Code that never leaves the company
Software development automation startup Factory revealed on the 11th how it runs its own development agent on NVIDIA DGX Spark. In a post on its official X (formerly Twitter) account, the company said, "Security-sensitive organizations should be able to run a software factory without moving source code or execution environments outside their walls." The setup, it said, offers a path to handle everything from identifying issues to fixing code entirely on local hardware.
Factory builds agentic development tools that automate work spanning understanding a codebase, investigating issues, and making fixes. What stands out in this announcement is the infrastructure pairing: NVIDIA's compact AI supercomputer, DGX Spark, combined with NVIDIA's own open model, Nemotron 3.5 Lightning.
Why DGX Spark
DGX Spark is a desktop-sized AI computer aimed at organizations that want to run large language models locally without relying on data-center-scale GPUs. On the 8th, NVIDIA AI's official account introduced a guide for connecting multiple DGX Spark units to run local AI, reflecting growing demand from individual companies and teams to build their own AI infrastructure. Rather than sending code to the cloud, the trend is toward finishing the job on equipment inside the office.
Factory's pairing is a concrete example of this shift. In industries such as finance, defense, and healthcare, where regulations or security policies restrict code from leaving the premises, cloud-based AI coding tools are often unusable. A local factory running on DGX Spark could offer such organizations a practical alternative.
Why Nemotron 3.5 Lightning fits this pairing
The model Factory chose is NVIDIA's open model Nemotron 3.5 Lightning. It uses a hybrid Mixture of Experts (MoE) architecture, with a total parameter count of 30 billion but reportedly only around 3 billion active parameters used during actual inference. Because it activates only the parts it needs rather than the whole model, the load on local hardware is comparatively lighter.
The model received day-zero support from vLLM on the 10th, and was noted for boosting agentic task speed through multi-token prediction and speculative decoding techniques such as DFlash and DSpark. Its supported hardware range is also broad, spanning DGX Spark, Jetson, H100, and B300/GB300. It reportedly supports a context length of up to 1 million tokens, which is said to be advantageous for working with entire large codebases at once.
The setup at a glance
| Component | Role |
|---|---|
| Factory | Agentic development tool that automates code investigation and fixes |
| NVIDIA DGX Spark | Compact AI computer for running large models locally |
| Nemotron 3.5 Lightning | Hybrid MoE open model with 3 billion active parameters |
vLLM Adds Day-One Support for Nemotron 3.5 Lightning, Bundles Three Inference Accelerators
What actually changes
Until now, code automation tools have generally had to route through cloud APIs, meaning source code inevitably moved to and from external servers. Factory's announcement is an attempt to break that assumption. Organizations that couldn't adopt AI coding tools because of security policy now have room to handle the same kind of work in-house with just a DGX Spark and an open model. Still, how close performance comes to that of large cloud models, and what real-world adoption costs look like, are hard to gauge from this announcement alone. What does seem clear is that competition over local AI infrastructure is now playing out simultaneously on both the hardware and open-model fronts.
