v1.7.3

🚀 Added

🍎 RF-DETR finally feels at home on a Mac

Until now, RF-DETR on a Mac ran slower on Core ML than on plain CPU: ONNX Runtime picked the legacy Core ML format, which can't do LayerNorm or GELU, and chopped the graph into ~100 pieces. Two fixes. ONNX models now ask for an ML Program, so RF-DETR Nano goes from 72 ms to 16 ms end to end. And when a model has a native Core ML package (Roboflow Train has built one for every RF-DETR since July 2025), AutoModel runs it directly: Nano 10 ms, Large 19 ms, versus 49 / 200 ms on CPU, with COCO AP within 0.01. Author's numbers, M4 Max. Oversized models are now rejected before the expensive compile, not after (@yeldarby, #3082, #3085; @dkosowski87, #3100).

🧬 Your classifier already knows what "normal" looks like

New roboflow_core/embedding_model@v1 pulls the feature vector (or raw logits) out of a classifier you already trained: ResNet, ViT or DINOv3, plus the pretrained ResNet aliases. No retraining. Plug the output straight into Cosine Similarity for novelty detection or "find me more like this". Also available as POST /infer/embeddings and in the SDK (@yeldarby, #3095).

🧠 Two more brains

Claude Sonnet 5.5 in anthropic_claude@v5 and GPT-6.1 Sol in open_ai@v7, with low to max reasoning effort (@SkalskiP #3076, #3081).

🔧 Fixed

• A long CLIP prompt is your problem now, not ours — text past CLIP's 77 tokens returns 400 instead of 500, on both model backends and in the CLIP Workflow blocks (@PawelPeczek-Roboflow, #3084).
• RF-DETR trusts the model over the paperwork — some packages said 640×640 while the ONNX file wanted 384×384, so every call failed. The loader now believes the weights (@yeldarby, #3089).
• Your webcam does 30 FPS again — fourcc is applied before fps, so a Logitech C920 at 1080p switches to MJPG instead of crawling at 5 FPS in YUYV (@ColePBryan, #3105).
• RF-DETR works on a read-only Jetson 6.2 container — Triton's kernel cache now goes to /tmp, as on JetPack 7.2 (@JeremyGracey-AI, #3072).
• The community forum link goes to the community forum (@pratikgx, #3096).

⚙️ Execution Engine v1.16.1

1.16.0 → 1.16.1: dependency preloading tells classification, feature-vector and logits uses of the same model apart. Existing workflows need no migration. Details in workflows/CHANGELOG.md.

🚧 Maintenance

• Versions — inference 1.7.3, inference-models 0.39.1, roboflow-workflows 0.2.4, streamvision 0.1.0 (@grzegorz-roboflow, #3114).
• supervision 0.30.6 — results stay the same, with two small exceptions: Time in Zone can flip for an anchor within 1 px of a zone edge, and Detections Stitch with NMM is slower on large frames in NumPy mode (@grzegorz-roboflow, #3079).
• Security refresh — Python and npm dependency patches (#3111, #3108).
• CI — automated PR review runs on Claude Opus 5.5 (#3107), the fast-track lane dropped a dormant suite (#3077), integration tests cache their assets (@ecarrara, #3104), linters are happy (#3070).
• Docs site analytics move off Segment (@franklin-robo, #3071).

New Contributors

• @franklin-robo made their first contribution in #3071
• @pratikgx made their first contribution in #3096
• @JeremyGracey-AI made their first contribution in #3072
• @ColePBryan made their first contribution in #3105

Full Changelog: v1.7.2...v1.7.3