$ cat wiki/models/lfm3-1-7b.md
LFM3-1.7B
Spec
| Attribute | Value |
|---|---|
| Developer | Liquid AI |
| Released | 2026-10-27 |
| Announced | 2026-10-27 |
| Context window | 256K tokens |
| Pricing | unknown |
| License | unknown |
| Availability | Open weights on Hugging Face; Liquid AI Playground and API; local deployment |
| These details come from the launch announcement. Open-weight availability does not establish license terms; the linked license and model card were not reviewed. (source) |
Release Date
Released 2026-10-27 as the text model in the first LFM3 family. This record covers LFM3-1.7B, not the separate vision, audio, omni or decision variants. (source)
Benchmarks
Liquid reports 45.60 on LiveClaw and 43.40 on SWE-bench Verified using OpenCode. Its comparison gives Qwen3.5-9B 35.77 and 51.20, respectively: the ordering changes with the task. These are vendor-reported scores, kept in the table's units rather than relabeled as percentages. (source)
Use Cases
The model targets local Agents (LLM Agents) and specialized applications. Its post-training includes supervised fine-tuning, mixed reinforcement learning, distillation from specialist teachers and agentic RL in real harnesses. The backbone combines gated short convolutions, sliding-window attention and global attention; Liquid says models can be adapted on a single GPU. (source)
Compared To
The launch's numerical predecessor is LFM2.5-2.6B. The table gives it 14.70 on LiveClaw and 18.00 on SWE-bench Verified with OpenCode. These comparisons retain the publisher's task labels and do not establish performance in a different harness. (source)
Sources
- Liquid AI launch, including the text-model benchmark table. (source)