Is LFM3-1.7B state of the art?Pending benchmarks.Announced Oct 27, 2026; no independent results yet.
Liquid AI's LFM3 text model for local agents and specialized applications, on a new tri-hybrid architecture of gated short convolutions, sliding-window attention and global attention, with a 256K-token context. Open weights under the LFM Open License v1.0, with the pre-trained LFM3-1.7B-Base released alongside.
The details so far
- Lab
- Liquid AI
- Announced
- Oct 27, 2026
- API
- Live · Liquid AI API and Playground, per Liquid's blog
- Open weights
- Published
- Parameters
- 1.73B
- Architecture
- 24 layers: 13 gated short convolutions, 5 sliding-window attention (256-token window), 6 global attention
- Context
- 256K tokens (input + output)
- Training
- 22T tokens of pre-training (last 3.5T curated mid-training with agentic traces); SFT, mixed RL, multi-domain on-policy distillation and agentic RL
- Languages
- English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian, Polish
- Memory
- About 1 GB at 4-bit precision, per Liquid
- Checkpoints
- Post-trained LFM3-1.7B and pre-trained LFM3-1.7B-Base on Hugging Face (plus GGUF, MLX and ONNX builds)
- License
- LFM Open License v1.0
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: outperforms LFM2.5-2.6B with 35% fewer parameters, averaging 49.37 on Liquid's agent, tool-use, coding, instruction-following, knowledge and long-context table vs 47.24 for Qwen3.5-9B and 39.27 for Qwen3.5-4B.
Lab claim: at 32K context (llama.cpp, Q4_0) it prefills at least 2.5x and decodes at least 1.8x faster than MiniCPM5-2B, Qwen3.5-4B, Qwen3.5-9B and Gemma 4 E4B on three devices, using the least memory.
Sources
LABLiquid AI: Introducing LFM3liquid.ai/blog/lfm3 · Oct 27, 2026LABLiquid AI announcement on Xx.com/liquidai/status/2115128329003303107 · Oct 27, 2026REPOLiquidAI/LFM3-1.7B on Hugging Facehuggingface.co/LiquidAI/LFM3-1.7BREPOLiquidAI/LFM3-1.7B-Base on Hugging Facehuggingface.co/LiquidAI/LFM3-1.7B-BaseTracking continues until an independent board lists it.