Blog & deployment guides

TCP for server-to-GPU communication
June 19, 2026
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.
Multimodal processor rewrite: 5x lower latency
May 13, 2026
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
May 12, 2026
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
April 2, 2026
Google is back on the open model map with Gemma 4. Four multimodal models, released today, now available on Overshoot for real-time inference.
Deploying VLMs through vLLM: Field Guide
February 5, 2026
Reproducible deployment guides for 30 vision language models across 11 families: setup instructions, deploy commands, troubleshooting, and benchmarks.

Deployment guides

Qwen3.5Qwen3.5-2B Qwen3.5-4B Qwen3.5-9B Qwen3.5-27B Qwen3.5-35B-A3B

Qwen3-VLQwen3-VL-2B Qwen3-VL-4B Qwen3-VL-8B Qwen3-VL-30B-A3B Qwen3-VL-32B

Qwen2.5-VLQwen2.5-VL-3B Qwen2.5-VL-72B

InternVL3.5InternVL3.5-4B InternVL3.5-8B

InternVL3InternVL3-2B InternVL3-8B

Molmo2Molmo2-4B Molmo2-8B Molmo2-O-7B Molmo2-VideoPoint-4B

MiniCPM-VMiniCPM-V-4.5