Self-hosted Qwen3.8-27B on one H200 reached 84% of Gemini's score. Most of its gap is picking the right object, not drawing the box.And Gemini 3.8 vs 3.7: same accuracy, half the time, 1.8× the cost.
Upgrading Gemini 3.7 → 3.8 moved our video grounding score by −0.4 points.Changing what one pipeline stage tells the next about "which object": +4.0.Fix the pipeline before you upgrade the model. 🧵
Astra: $105 per hour of video. Gemini: $3.70.Pay for Astra on high-stakes audits. Use Gemini for 24/7 cameras.Full benchmark, with Claude Opus 5.5 and Qwen3.8-27B: blog.agnify.ai/gpt-6-astra-gr…
The twist: on its first run Astra only tied Gemini.It kept rewording object names, and our pipeline split one object into two tracks. We switched to referring to objects by number, and its score jumped to 0.622.A smarter model doesn't fix a pipeline that fights it.
Everyone is chasing bigger VLMs.
Production systems are moving in the opposite direction.
Fast, specialized models handle most of the workload. Foundation models step in only when reasoning is required.
That’s why we extended D-FINE-seg with semantic segmentation. One production-ready framework now covers detection, instance segmentation, and semantic segmentation while remaining fast and efficient.
Read more: blog.agnify.ai/d-fine-seg-sem…
46 Followers 118 FollowingCameras are the eyes of AI. Building Agnify to turn video into intelligence. Former co-founder & CTO of Avride. Previously Yandex
46 Followers 118 FollowingCameras are the eyes of AI. Building Agnify to turn video into intelligence. Former co-founder & CTO of Avride. Previously Yandex