LFM3
Liquid AI's ~2B-class open-weight family: a 1.7B text model with 256K context that fits in about 1 GB at 4-bit.
Overview
Liquid AI released the LFM3 family on October 27, 2026: six ~2B-class models branched from LFM3-1.7B-Base, covering text, vision, audio and omni inputs plus two decision models. The text model, LFM3-1.7B, has a 256K context window and uses a 24-layer tri-hybrid design of 13 gated short convolutions, 5 sliding-window attention layers and 6 global attention layers. It was pre-trained on 22T tokens and post-trained with SFT, mixed RL, on-policy distillation and agentic RL. Liquid reports an average of 49.37 across its benchmark set versus 31.55 for LFM2.5-2.6B, with SWE-bench Verified at 43.40 and RULER 128K at 65.56; larger models still lead on agentic coding and knowledge.
This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.
Key features
- LFM3-1.7B: 256K context, about 1 GB at 4-bit precision
- Tri-hybrid layers: 13 gated short convolutions, 5 sliding-window and 6 global attention
- 22T-token pre-training; four-stage post-training including agentic RL
- Self-reported IFEval 96.66, BFCL V4 71.39, SWE-bench Verified 43.40
- KV cache of 1.6 GB at 128K context versus 4.3 GB for Qwen3.5-4B
- Variants for vision, audio and omni input; fine-tunable on a single GPU
Use cases
- On-device assistants and tool-calling agents on phones and laptops
- Long-context document processing within tight memory budgets
- Task-specific fine-tuning on a single GPU
Pricing
Open weights on Hugging Face; also on the Liquid AI Playground and API (API pricing and license terms not stated on the launch post).
Pricing reflects the source record and may change. Confirm current terms with the provider.
Dates and sources
- Released
- 27 Oct 2026
- Record updated
- 28 Oct 2026
- Record last checked
- 28 Oct 2026