Open VLMs reach commercial viability for egocentric data processing
kuaythrone · reddit · 2026-08-31
Evaluated the latest open-weights VLMs for egocentric data processing using HFlow, based on Build AI's Egocentric-10k evaluation.
Agreement with Gemini 2.5 Flash baseline:
- Gemini 2.5 Flash: 91.65%
- GLM 5.3 Flash: 91.00%
- Gemma 4 26B-A4B: 90.87%
- Qwen 3.8 27B: 90.79%
- Inkling Small: 85.21%
Key Takeaways:
- Gemma 4 stands out with performance on par with Gemini at 19x lower cost.
- Both Gemma and Qwen are practical for self-hosting, enabling private processing without data egress.
- Modern open-weights VLMs are becoming sufficient for large-scale egocentric data processing, with differentiators shifting to cost, throughput, reliability, and ease of self-hosting.
More from Models
- OpenAI stealth-testing mysterious internal model 'Aeon' in Codex desktop app and Dots — cedric_chee · 2026-10-16
- Argon edges out astra as daily driver, but Claude still wins heavy coding work — iruletheworldmo · 2026-10-16
- NVIDIA ships FP4-quantized Qwen3.8-Flash for local use, trending on Hugging Face — nvidia · 2026-10-16
- RoboVista benchmark: Gemini-4 Argon hits 81.9% on robotics understanding — berkeley_ai · 2026-10-16
- Dev reports decision model jev beats gpt-6-luna on quality in real training pipeline — genmon · 2026-10-16
- Google won't publish Argon 4 limits in AI Pro—users compare experiences — Consistent-Text-2662 · 2026-10-16