Blog & deployment guides
All You Need to Know about JEPA
JEPA predicts what a missing region represents instead of reconstructing its pixels — and how I-JEPA learns visual structure in latent space.
The Real World Is Not a Dropdown Menu: Why Fixed-Label Vision Models Break Down
Why fixed-label vision models break down in the open world, and how CLIP-style zero-shot learning and modern VLMs change the interface.
From Pixels to Tokens: How Vision Transformers Read Images
How Vision Transformers turn spatial images into patch tokens, why pixels are too costly, and how visual tokens connect to language models.
VLMs Can See Objects. But Why Can't They Count Them?
Counting looks simple for VLMs but fails under clutter, composition, and scale. How divide-and-conquer pipelines like LVLM-Count improve accuracy.
When VLMs Answer Without Seeing: The Mirage Problem
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.
An Introduction To Token Pruning For VLMs
Why vision language models can drop visual tokens without losing understanding, from parameter pruning to PuMer and FastV and the efficiency tradeoff.
TCP for server-to-GPU communication
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.
Can VLMs Understand Without Generating?
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: Papers on Real-Time Vision
CVPR 2026 accepted about 4,000 papers; roughly 50 target real-time vision. Our shortlist of seven spans streaming compression, inference, and reasoning.
Multimodal processor rewrite: 5x lower latency
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.
Gemma 4: a 10x vision speedup from one fix
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.
Day Zero Gemma 4 Support on Overshoot
Google is back on the open model map with Gemma 4. Four multimodal models, released today, now available on Overshoot for real-time inference.
Open Source Vision Language Models (2026)
A catalog of every significant open-source vision language model released since December 2024: 56+ models across 13 families, with deployment support.
Qwen 3.5: Architecture and Model Selection
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.
Deploying VLMs through vLLM: Field Guide
Reproducible deployment guides for 30 vision language models across 11 families: setup instructions, deploy commands, troubleshooting, and benchmarks.