Quick answer
LFM2.5-2.6B is Liquid AI’s 2.69B-parameter dense, text-only model for agentic and on-device workloads. Its official model card lists a 131,072-token context window, native tool calling, and native, GGUF, MLX, and ONNX artifacts.
It is a strong candidate to evaluate when your agent needs bounded tool use, data extraction, RAG, or long tool traces on constrained hardware. It is not a blanket replacement for a cloud model: Liquid AI does not recommend it for agentic coding or knowledge-heavy tasks, and maximum context does not prove acceptable memory use or latency on your device.
Where the model fits
| Workload | Why it may fit | Acceptance check |
|---|---|---|
| Bounded tool workflow | Native function-calling format | Correct tool, arguments, permission, and final answer |
| Data extraction | Small local text model and structured workflow | Field accuracy, abstention, malformed-input handling |
| Retrieval-augmented answer | Local generation with supplied context | Citation support, conflict handling, no unsupported answer |
| Long multi-step trace | 128K published maximum context | Real memory, latency, retrieval quality, and trace pruning |
| Agentic coding | Provider says not recommended | Choose another model or constrain to reviewed assistance |
The model’s release record supports a candidate evaluation, not a production conclusion. Pick representative tasks and compare it with the current baseline on completed-task quality, not parameter count alone.
Build a safe evaluation
- Pin the exact repository, revision, format, quantization, runtime, and prompt template.
- Separate read-only tools from tools that write, send, buy, delete, or change access.
- Validate every tool name and argument outside the model.
- Require explicit approval for material actions and scope changes.
- Replay versioned success, ambiguity, malformed input, denial, timeout, and retry cases.
- Score task completion, incorrect calls, false success, refusal quality, recovery, latency, and peak memory.
- Keep an independent stop, audit, and rollback path.
On-device inference does not automatically keep the whole workflow local. Retrieval services, tool APIs, telemetry, synced files, and model downloads remain separate data paths.
Choose a deployment path
Use the local setup guide for a reproducible first run, the runtime comparison for format selection, and the device sizing guide before making a phone or laptop fit claim. The implementation planner keeps those decisions together.
Frequently asked questions
What is LFM2.5-2.6B designed for?
Liquid AI describes LFM2.5-2.6B as a 2.69B-parameter dense, text-only model post-trained for agentic workloads, native tool calling, long-context workflows, data extraction, and RAG.
Is LFM2.5-2.6B suitable for agentic coding?
Liquid AI’s model card explicitly does not recommend it for agentic coding or knowledge-heavy tasks. Test a different model or a tightly bounded assisted workflow for those jobs.
Does on-device deployment make an agent private or safe?
No. Local inference can reduce where model inputs travel, but tool calls, logs, retrieval sources, crash reports, permissions, updates, and synced outputs each have separate data and safety boundaries.
Official sources
Source check: August 5, 2026. Provider benchmarks, device results, and maximum context are starting evidence. Verify the exact artifact, runtime, license, hardware, data path, and task behavior independently.