Fixstars Launches Contract Service for Porting and Accelerating AI Models for Embedded Devices in One-Third the Conventional Development Time

Sep 28, 2026 – Fixstars Corporation (TSE Prime: 3687, US Headquarters: Irvine, CA), a leading company in performance engineering technology, today announced the launch of the “1/3 Delivery Time: AI Model Porting & Acceleration for Embedded Devices,” a contract development service that accelerates AI models and software running on embedded devices and ports them to a wide range of devices in significantly less time than before. The service was made possible by secure AI agents that Fixstars has developed and operates in-house, which dramatically shorten the development time required for acceleration and porting.

In recent years, there has been growing demand to port compute-intensive workloads such as AI inference and image and signal processing to edge devices and automotive SoCs. However, optimizing these workloads requires specialized expertise and extensive verification effort, which often prolongs development.

This service builds on the software acceleration and porting expertise Fixstars has cultivated over more than 20 years, combined with agentic engineering on “Fixstars Vega,” the company’s in-house AI platform. By leveraging the latest open-weight LLMs, it significantly shortens development time while maintaining confidentiality, quality, and performance.

Key Features of the “1/3 Delivery Time: AI Model Porting & Acceleration for Embedded Devices”

Feature 1: Shorter schedule: development time cut to about one-third

Development time for acceleration and porting is reduced to approximately one-third of the conventional timeframe. For example, a project that would previously have taken six months is expected to be delivered in about two months. Shorter development cycles enable faster time-to-market and an earlier start on the next development project.

Feature 2: Standard pricing: no rush fees

No rush fees apply despite the shorter lead time. Development costs remain on par with conventional contract development.

Feature 3: Better quality and performance

Comprehensive testing ensures quality, and performance is pushed to the limits of the hardware. Deliverables include test results, benchmark results, and design documentation.

Feature 4: Rigorous confidentiality

The AI agents used in development run on an AI platform operated in-house by Fixstars. Specifications, source code, and data entrusted by customers are never sent to external AI services, nor are they used to train external AI models.

Note: Actual reductions in time and cost vary depending on the development scope and will be presented individually following a free assessment.

Service Overview

The service covers the acceleration of AI models and software, as well as porting to a variety of embedded devices. In both cases, Fixstars provides end-to-end support, from performance verification on actual hardware to continuous improvement after delivery.

Services Offered

– Performance optimization: Increases processing speed several-fold on existing systems. Based on profiling, techniques such as multithreading, GPU utilization, and quantization are applied. Target workloads include compute-intensive processing in general, such as image and signal processing, AI inference and training, simulation, and numerical computation.

– Porting to embedded devices: Ports AI models and applications to target hardware, maximizing hardware performance while maintaining accuracy. Supported targets include NVIDIA DRIVE / Jetson, Renesas R-Car, Qualcomm Snapdragon, Tenstorrent, and various FPGAs and DSPs.

Service Process

1. Free assessment: Review the target workload, current performance, and target environment, and estimate the achievable reduction in development time and attainable performance.

2. Quotation: Present the development period, quality criteria (test policy and performance targets), and costs; work begins upon agreement.

3. Development: Engineers and AI agents work collaboratively, with progress and deliverables shared weekly.

4. Delivery: In addition to working software, deliver test results, benchmarks, and design documentation.

Consultations and requests for processors or workloads outside the listed scope are also accepted on a per-project basis.

Details: https://www.fixstars.com/en/products-services/embedded-software-development
Inquiries: https://www.fixstars.com/en/contact

How Short Lead Times Are Achieved: Fixstars Vega and Agentic Engineering

Fixstars engineers have adopted and practice agentic engineering, an approach in which engineers develop software in collaboration with AI agents.

As an environment for agentic engineering, Fixstars has built “Fixstars Vega,” its in-house AI platform, and developed a performance engineering harness^[1] (an AI agent execution environment specialized for acceleration and porting) that runs on it and is continuously improved.

Fixstars Vega is an AI platform built and operated so that AI processing is completed entirely within the company’s internal infrastructure and network using open-weight LLMs^[2]. The harness incorporates embedded software development and performance engineering know-how that Fixstars has accumulated over many years, including chip-specific optimization patterns, quantization strategies, and key points for performance measurement. This enables highly specialized development, unattainable with general-purpose AI tools, to be delivered with high quality and short lead times.

• Dedicated acceleration and porting specialists develop and improve the harness daily, continuously raising the development efficiency of all engineers.
• Engineers handle design, technical decisions, and code reviews, combining the speed of AI with professional quality.
• The harness embodies Fixstars’ own technical know-how; customers’ confidential information is never used for AI training.

Ultra-Short Lead Time Case Study: Accelerating Inference of an Autonomous Driving AI Model

For an AI inference optimization task originally estimated to require approximately two person-months, an engineer using AI agents achieved the target speed in just one week.

This represents a reduction in effort of approximately one-eighth, the largest achieved in the company’s internal verification.

• Development effort: 1 person-week (manual: approx. 2 person-months) ^*1
• Task: Inference acceleration of a 3D object detection AI model for autonomous driving
• Inference speedup: 3.95× (162.3 → 41 ms/iteration)
• Approach: TensorRT conversion, PTQ^[3] (weight quantization)^*2, implementation of custom CUDA kernels as TensorRT plugins, etc.

*1 The development period reflects results achieved by an engineer experienced in deploying this model.
*2 Accuracy changed from mAP 0.428 to 0.417 (a slight decrease due to PTQ, expected to be recovered through calibration; manual optimization maintained 0.428).

[1] Harness: An environment for running AI agents. Here, it refers to Fixstars’ proprietary AI agent execution environment incorporating the tools and optimization know-how required for acceleration and porting.

[2] Open-weight LLM: A large language model whose weights (parameters) are publicly released and can be operated in a company’s own environment. Inference can be run on in-house hardware without relying on external services.

[3] PTQ (Post-Training Quantization): A technique that quantizes the weights of a trained model to lower bit widths while minimizing the impact on accuracy, making processing faster and lighter.

About Fixstars Corporation

Fixstars is a technology company dedicated to accelerating AI inference and training through advanced software optimization solutions. It supports innovation in healthcare, manufacturing, finance, mobility, and other industries. For more information, visit: https://www.fixstars.com/