D

Decart Optimization Stack

Infrastructure service

by Decart

The Decart Optimization Stack (DOS) is Decart's inference and training optimization infrastructure, spanning hardware-aware model design, kernel tooling, proprietary compilers and inference optimization. It is sold to hardware providers and AI teams as engagements covering benchmark optimization, customer-defined kernel and compiler work, cross-workload efficiency gains, and profiler and simulator licensing, and is marketed as hardware-agnostic across GPUs, TPUs, Trainium and AMD accelerators.

Find alternatives to Decart Optimization Stack

No public pricing. The page's only CTA is 'Contact us', to scope either 'a scoped milestone-based pilot' or 'a long-term strategic partnership' - that contact route is how it is bought on its own, separately from the Decart API Platform's per-second rate card. One named offering is 'Profiler and simulator licensing'.

What it does

• Model inference
• Model training
• GPU programming

Sources

Pricing

Not found · 5 Sep 2026

No pricing figures on the page; only a 'Contact us' CTA for a pilot or partnership.

Description

Company disclosure · 5 Sep 2026

H1 'The ultra-optimized infrastructure for AI'; 'The Decart Optimization Stack (DOS) squeezes every ounce of performance from every chip, across inference, training and hardware...'. Offer bullets include 'Gold-standard benchmark optimization for low-latency AI workloads', 'Customer-defined optimization on workloads and success metrics', 'Cross-workload efficiency gains across inference and training', 'Profiler and simulator licensing'. Independently re-fetched by the verification pass 2026-09-05. NOTE the live path is /optimization-engine; /optimization is a 404.

Sold within

Official documentation · 15 Sep 2026

decart.ai/optimization-engine, the vendor's own page — recorded as official-docs rather than the more specific company-disclosure because validate-data.py's standalone+null-parent check accepts only pricing-page/official-docs/not-found for this field; presented as its own line of business separate from Decart's hosted service.

Deployment

Official documentation · 5 Sep 2026

LOW CONFIDENCE, inferred. Marketed as running on the customer's own compute - 'Extract more performance from the compute you already have', 'Run optimized workloads across every major chip architecture' (GPUs, TPUs, Trainium, AMD). The page never uses the words cloud or on-prem.

Also from Decart

• Decart API Platform Model API

  Hosted API for Decart's real-time video and world models.

• Cogito Model API

  Cogito is Decart's OpenAI-compatible LLM inference API, serving frontier open-weight models including Kimi K2.6, Kimi K2.7 Code, GLM-5.2, Qwen3 235B and GPT-OSS 120B across Trainium, TPU and GPU capacity. A self-serve Standard tier is billed per token, while an ultra-fast reserved tier advertises 1,000+ tokens per second. It is built on Decart's own DOS optimization stack.

• Delulu (by Decart) Application

  Delulu is Decart's consumer mobile app for AI photo transformation, published on iOS and Android by Decart.AI, Inc. Its store listing describes it as a way to 'Turn any photo into something fun, weird, or just really good-looking, with a single tap', browsing preset styles and creating stickers to share.

• Lucy Application

  Real-time video editing platform that edits and transforms live video at streaming speed, for enterprises, brands, creators and live/streaming/gaming use cases; built on Decart's DOS infrastructure.

• Oasis 3 API service

  Interactive world model for Physical AI, generating realistic, controllable real-time simulation environments for training autonomous systems, distributed as an API/SDK product built on Decart's Optimization Stack (DOS) infrastructure.

Open Decart in the directory

Something wrong here? Send a correction — quote this product id: decart-dos.