HPT

HyperGAI's pitch is narrow and honest: small multimodal models you run yourself. The interesting release is HPT 1.5 Edge, a ~4B vision-language model the team positions at the front of its size class for edge and mobile hardware. That is a different buying conversation from a hosted API — you are evaluating whether a small open model clears your accuracy bar on your own images, then deciding if you can serve it. We'd reach for HPT 1.5 Air when you want a stronger reasoning floor and 8B on a Llama 3 base is acceptable. HyperGAI says it beat bigger proprietary models on some benchmarks, which is worth verifying on your task rather than taking at face value. When to pick this: on-device visual Q&A, private document parsing, chart and figure extraction, or code generation from screenshots where sending images to a third party is a non-starter. Fine-tuning scenarios suit it too — you can adapt the weights to a niche visual domain instead of prompt-engineering around a closed model's gaps. When to pass: if you need a chat product tomorrow, this is a research release, not an app. There is no hosted inference endpoint in the sources, and generation of images or video is a stated roadmap item rather than something you can download. Non-technical teams will bounce off the setup cost. In practice the caveats are the usual open-weight ones — you build the serving layer, you own latency and GPU sizing, and you benchmark against your own data because vendor benchmark charts rarely match production inputs. Budget evaluation time before you commit. The closest alternative is an open LLaVA-class multimodal model, or a hosted vision API if data-residency rules let you use one. Where it bites is the boundary between Edge and Air: pick Edge for constrained hardware, Air when you have