LiquidAI releases 3B local VLM: Beats Gemma-4 in performance
Designer_Athlete7286 · reddit · 2026-08-15
LiquidAI launched LFM2.5-VL-3B, a 3.1B parameter local vision-language model. It runs on llama.cpp, MLX, vLLM, and WebGPU. Benchmarks show it outperforms Gemma-4 E4B (8B) and edges out Qwen3.5-4B, with a significant jump in screen understanding. This shifts simple document/screen tasks from recurring API costs to a one-time engineering effort with local data privacy.
More from Infra
- NVIDIA open-sources NeMo Switchyard for dynamic model routing in agent workflows — NVIDIAAI · 2026-08-15
- Vercel ranked as the world's fastest AI Gateway infrastructure — cramforce · 2026-08-15
- Qwen3.8-2.4T-A95B deployment guide: NVFP4 needs 8×B300, TP must divide 64 — Necessary_Gazelle211 · 2026-08-15
- mcpp: Auto-generate MCP servers from C++ code via reflection — karurochari · 2026-08-15
- RTX 3090 gets 35 t/s on Qwen 3.8 27B — cviperr33 · 2026-08-15
- CME to launch futures contracts tracking Nvidia H100/B100 compute costs — AccBalanced · 2026-08-15