When the open model wins, the open model ships.
Qwen, Llama, Mistral, DeepSeek, Kimi, GLM, MiniMax and Hermes, served on Ollama and pulled from Hugging Face — plus Grok and Gemini benchmarked alongside them. VeUP holds no partnership with any of them, which is exactly why this is the most honest column in the practice: a model gets the workload only by beating the frontier labs on the customer's own data.
The engagements whose records name these models.
The flagship exhibit is a workload an open-weight model won outright against Claude Opus, Llama 4 Maverick, Llama Guard 4 and Nova Premier — scored on an identical labelled dataset, every candidate invoked through Amazon Bedrock.
Qwen 3 VL is the sole production engine, selected on the numbers after Llama 4 Maverick, Llama Guard 4, Nova Premier and Claude Opus were scored on the same labelled dataset. 54% → 99.58% NSFW accuracy with prompt engineering alone — no fine-tuning — at roughly $1 per 1,000 images.
The serving pattern open-weight models land on: GPU-served moderation models on dedicated Amazon EKS node groups behind NVIDIA Triton, migrated off GCP with zero downtime.
Eleven model providers in production — including self-hosted open weights.
Not a customer claim: this is VeUP's internal engineering estate, the machine that produced this proof library. It is the reason open-weight serving is a practised skill here rather than a slide.
- Self-hosted Qwen, Kimi, GLM and MiniMax alongside Claude and GPT in the internal fleet.
- One governed tool server — 350+ tools across 14 business systems — model-agnostic by construction.
- GPU serving on Amazon EKS behind NVIDIA Triton, the same pattern shipped for customers.
- Every call metered through one gateway, so a model swap is a config change with a receipt.
15 accounts where Open-weight is on the table.
These are customer accounts VeUP runs where this model family came up on a recorded customer call — architecture sessions, evaluations, roadmap conversations. That is engagement evidence, not a statement that the customer runs these models in production. Promotion to a production claim means a published case study, which is what the section above holds. 5 of these 15 are named here because they already have one. Every other row stays anonymized — including customers who do have a published case study but have not agreed to be named, who appear by description only.
- Identity held on fileTelecommunicationsBuildIgniteSellOperate28 AWS accounts
- Identity held on fileInformation Technology and ServicesSellOperate11 AWS accounts
- BrontoSoftware DevelopmentSellOperate10 AWS accounts
- Identity held on fileComputer SoftwareBuildSellOperate9 AWS accounts
- Cluster Systems (product: ClusterPOS)Software DevelopmentBuildIgniteOperate9 AWS accounts
- Identity held on fileConstructionSellOperate8 AWS accounts
- An AI-powered frontline-workforce enablement platformInformation Technology and ServicesSellOperate6 AWS accounts
- A consumer connected-home-security IoT platformComputer & Network SecuritySellOperate2 AWS accounts
- TenderdComputer SoftwareSellOperate1 AWS accounts
- Identity held on fileManagement ConsultingSellOperate0 AWS accounts
- MyCena Security Solutions—Sell0 AWS accounts
- Identity held on fileHospital & Health CareBuild0 AWS accounts
- A travel-technology company building a governed traveler-data platformLeisure, Travel & TourismBuildIgnite0 AWS accounts
- Identity held on file—Build0 AWS accounts
- ZenhubProgram DevelopmentBuild0 AWS accounts
Every account here is a VeUP customer where this family is in the work. The chips are the engagement motions VeUP runs at that account — Build means VeUP builds and operates there, Ignite that VeUP funds and resells the estate, Sell and Marketplace the co-sell and listing motions. Operate marks the estates where VeUP is the AWS payer of record. Accounts are resolved from the customer registry and the meeting archive on participant email domain. Counting conversations is not counting deployments — the number is here as scope, never as a claim. Linked AWS accounts describe estate size, not spend.
