Blog & deployment guides

Counting looks simple for VLMs but fails under clutter, composition, and scale. How divide-and-conquer pipelines like LVLM-Count improve accuracy.

Vision-language models can answer with confidence even when the image is blank, unrelated, or uninformative. An overview of mirages, why they matter, and how grounding checks can catch them.

Why vision language models can drop visual tokens without losing understanding, from parameter pruning to PuMer and FastV and the efficiency tradeoff.

Not every weight in a neural network is equally necessary. Weight pruning, structured pruning, and MoPE-CLIP for vision-language models.

A 5 MB upload over a 31 ms RTT path took 325 ms cold. Connection reuse and congestion window tuning cut it to 72 ms, explained with interactive simulations.

Vision language models have structural limits on visual understanding. We ask whether the strengths of generative models can transfer and help close the gap.

CVPR 2026 accepted about 4,000 papers; roughly 50 target real-time vision. Our shortlist of seven spans streaming compression, inference, and reasoning.

CPU preprocessing cost 428 ms p90 for a 15-frame Qwen video before it touched the GPU. We rewrote the processor bit-identical at 81 ms p90, a 5x cut.

A simple fix in the vLLM video encoder path batches the serial vision tower and speeds up Gemma 4 on vision tasks by 10x, while preserving bit identity.

Google is back on the open model map with Gemma 4. Four multimodal models, released today, now available on Overshoot for real-time inference.

A catalog of every significant open-source vision language model released since December 2024: 56+ models across 13 families, with deployment support.

A technical look at the Qwen 3.5 family: five natively multimodal models from 2B to 35B, hybrid attention, benchmark analysis, and which model to pick.

Reproducible deployment guides for 30 vision language models across 11 families: setup instructions, deploy commands, troubleshooting, and benchmarks.