LiquidAI Launches LFM2.5-VL-3B for On-Device Multimodal
pmttyji · reddit · 2026-08-12
LiquidAI announced LFM2.5-VL-3B, a hybrid multimodal model designed specifically for on-device deployment. It uses the 2.6B parameter LFM2.5 language model as its backbone, combined with a SigLIP2 NaFlex vision encoder to process both text and images.
Key Features & Performance:
- Improved Capabilities: Optimized for natural language grounding/object detection and full-page OCR with layout annotations.
- Efficient Edge Inference: Achieves 228 tok/s on Apple M5 Max and 116 tok/s on AMD Ryzen AI Max+ 395, fitting in under 3.3 GB of memory. It also reaches 20 tok/s on a Galaxy S26 Ultra.
- GPU Throughput: Reaches approximately 11K tokens/s (nearly 1B tokens/day) on a single NVIDIA H100 using vLLM at high concurrency.
Supporting a 32k context length and multiple languages, the model is recommended for single-turn, high-throughput, low-latency tasks (e.g., real-time automotive detection, document OCR, on-device translation). It is not recommended for long-context reasoning or highly technical blueprint Q&A.
Related event: LiquidAI Releases LFM2.5-VL-3B Edge Multimodal Model(4 posts)→
More from Models
- DeepSeek V4-Pro-0813 Rolling Out to API — strangedell123 · 2026-08-12
- DeepSeek V4-Pro Benchmarks Leaked, Approaching Top-Tier Models — MagicZhang · 2026-08-12
- Polymarket Odds: OpenAI Favored to Hold #1 AI Model by 2026, xAI Tied with Alibaba — Polymarket · 2026-08-12
- Gemini Loses Market Share as ChatGPT and Claude Surge, Multi-Source Data Confirms — The Decoder · 2026-08-12
- Grok 4.6 Crushes Competitors on Value; Grok 4.7 to Integrate SpaceX Data — GavinSBaker · 2026-08-12
- xAI Officially Launches Grok 4.6: Major Performance Boost at the Same Price — xiaosun86 · 2026-08-12