Oleksandr Kuzmenko, EditorPublished 5 min readAI-assisted · editor-reviewed
Why it matters
You can run 256K-token agentic workflows and multimodal inference locally in under 1 GB of memory.
TL;DR
- 01LFM3 runs text models with 256K context in ~1 GB VRAM at 4-bit precision.
- 02Architecture combines short convolutions and gated attention for 25% faster 8K context inference.
- 03Single-GPU fine-tuning allows custom local deployment for vision, audio, and agent execution.
Key facts
- Parameters (Base/Text)
- 1.7B (uses 35% fewer params than LFM2.5)
- Context Window
- 256K tokens
- Memory Footprint (4-bit)
- ~1 GB
- WebArena-Verified Score
- 38.42 (vs 3.02 for LFM2.5)
- Audio Latency
- 50 ms time to first token
- Pretraining Volume
- 22T base tokens + 3.5T mid-training
Tri-Hybrid Layer Architecture
Liquid AI's LFM3-1.7B-Base operates with 24 layers split into 13 gated short convolutions (kernel size 7), 5 sliding-window attention (SWA) layers (256-token window), and 6 global attention layers. Global layers apply partial RoPE to the high-frequency half of head dimensions to preserve distance-agnostic matching. Sandwich normalization and head-wise sigmoid gated attention suppress activation spikes, enabling clean 4-bit quantization down to ~1 GB VRAM.
Extended Context and Agentic mid-training
Pretrained across 22T tokens, the base model underwent mid-training on 3.5T curated tokens rich in tool calls and agentic traces, extending the context window to 256K tokens. On WebArena-Verified, LFM3-1.7B scored 38.42 (up from 3.02 in LFM2.5-2.6B), outperforming Qwen3.5-9B on prolonged browser and enterprise tool workflows.
Multimodal Edge Lineup
- LFM3-VL-2B: Visual model accepting video feeds. Merges sampled frame pairs into temporal slices to halve sequence lengths, scoring 71.4 across 62 benchmarks.
- LFM3-Audio-2B: Fully causal audio encoder yielding 50 ms time-to-first-token, streaming speech decoding at 87x real-time on Apple M4 Max via LiquidCodec.
- Agent Framework Compatibility: Natively adapted for Hermes Agent, Pi, and OpenClaw harnesses.
Try it in 2 minutes
pip install "transformers>=5.21.0" torch torchvision pillowbash
✓ When to use
- Deploying autonomous tool-calling agents locally on edge devices, laptops, or Raspberry Pi 5.
- Building privacy-preserving speech-to-action or video understanding tools on single GPUs.
✕ When NOT to use
- High-complexity coding across multi-repo codebases where frontier reasoning models are required.
- Zero-shot abstract mathematical problem solving.
What to do today
- Install transformers>=5.21.0 to experiment with LFM3-1.7B and LFM3-VL-2B locally.
- Evaluate LFM3-1.7B inside Hermes Agent or OpenClaw harnesses for local agent execution.