Showcase

Built with AgentVUE.

Real projects, delivered by the same agents you get with a licence.

AgentVUE

The workstation that builds AgentVUE

AgentVUE itself — the editor, the websites, the AgentVUE Business Kit (ABK) — is specified, built and maintained by AgentVUE agents on every release. What you buy is the workflow that ships this product.

  • Windows
  • Linux
  • AgentCode
  • AgentWork
  • AgentVUE Business Kit (ABK)
  • ShortVUE
vllm-gfx906-mobydick logo — the Moby Dick whale

vllm-gfx906-mobydick

Bringing a 2018 GPU back to life with today's open-source inference

A fork of vLLM optimised for AMD gfx906 GPUs — Radeon VII, Instinct MI50 and MI60. Started from AgentVUE prototypes, the project ships ready-to-use ROCm Docker images: FlashAttention through Triton AMD, MTP speculative decoding, multi-GPU tensor parallelism. The result: the latest SOTA models — Qwen 27B in AWQ INT4, context up to 262,144 tokens — running at respectable speeds on hardware the ecosystem had abandoned.

  • vLLM
  • ROCm
  • AMD gfx906
  • Radeon VII
  • MI50 / MI60
  • FlashAttention
  • MTP

Transparency: active development on this fork is slowing down in favour of a new project, vllm-plus, targeting multi-GPU consumer NVIDIA Blackwell systems interconnected with InfiniBand and consumer NVIDIA Ampere GPUs — RTX 3090 included.

See the GitHub repository

Ship your own project with AgentVUE

One desktop app — a complete autonomous-agent workstation. Specify, build, review and maintain repositories and non-code work with the same workflow that delivered AgentVUE.

Download
Get AgentVUETour