Liquid AI releases LFM2.5-8B-A1B, a device-optimized 8B MoE model for on-device agents
Boston startup says the 128K-context, 38T-token LFM2.5 upgrade delivers reliable agentic behavior, fast tool calling, and open weights for phones, laptops, PCs, robots, and lightweight servers.
By Ryan Merket · Published
Why it matters
On-device models that can run full agent loops shrink latency, cut cloud costs, and unlock offline and privacy-sensitive use cases. An 8B MoE with 128K context broadens what fits on consumer hardware.
Liquid AI released LFM2.5-8B-A1B, a device-optimized model designed to power real-life applications on phones, laptops, PCs, robots, and fast, lightweight server-side use cases, per the company's X post.
Key specs and claims:
- 8B MoE, 1.5B active
- Expanded 128K context
- LFM2.5 flagship hybrid MoE architecture
- Trained on 38T tokens + large-scale RL
- Fast, reliable tool calling, "punching above its weight, comparable to models with up to 4x its size"
- Customizable on a single GPU for specialized tasks
- LFM2 open-weight license
https://x.com/liquidai/status/2060023455290974474
Liquid AI frames this as a step change from LFM2-8B-A1B, citing:
- Training toks: 12T -> 38T
- Context length: 32k -> 128k
- instruction following (IFEval): 79.44 -> 91.84
- IFBench: 26.00 -> 56.47
- Muti-IF: 58.54 -> 79.93
- Tau2Telecom: 13.60 -> 88.07
- BFCLv3: 45.07 -> 64.36
- BFCLv4: 25.52 -> 48.50
The company says this delivers reliable agentic behavior at 8B parameters. It describes the model as built for the full agentic loop on a single machine, able to chain tool calls across complex instructions with a fast dispatch loop (ask, propose, confirm, run, repeat), and with a doubled vocabulary for non-Latin language support. Liquid AI emphasizes no API keys and no data leaving the machine.
To demonstrate, LocalCowork, the startup's open-source desktop agent, now runs on LFM2.5-8B-A1B on a single laptop, using 67 tools across 13 MCP servers with well under a second per dispatch. Watch the 3-minute demo.
https://www.loom.com/share/bc3faf8befb643baae3434cde098e95e
Background: The Boston-headquartered company, founded by Ramin Hasani in 2023, builds foundation models using liquid neural networks. Per GetLatka, the startup reached $13.2M in revenue in 2025 with an 88-person team and has raised $287.6M across two rounds, including a $250M Series A at a $2.4B post-money valuation.
Sources: Liquid AI on X; 3-minute demo; GetLatka