Native
Transformers, vLLM, SGLang, or fine-tuning
Use the exact LiquidAI/LFM2.5-2.6B ID.
Liquid AI model record · Source checked August 5, 2026
LFM2.5-2.6B puts agent tool use into an on-device-sized model.
Liquid AI publishes a 2.69B dense, text-only checkpoint with 131,072-token context and native tool calling. It ships in native, GGUF, MLX, and ONNX formats. That record supports local evaluation—not a promise that every phone, harness, task, or Flowith account is ready for it.
Verified model record
Checkpoint
LiquidAI/LFM2.5-2.6B, post-trained for agentic workloads
Architecture
2.69B-parameter dense, text-only LFM2.5 model
Context
131,072 tokens in the official model card
Published formats
Native weights, GGUF, MLX, and ONNX
Recommended work
Tool use, data extraction, RAG, and long-context workflows
Published caution
Liquid AI does not recommend it for agentic coding or knowledge-heavy tasks
Flowith status
Not established by the provider release; verify the live workspace separately
Format decision
Choose the package from the target runtime
GGUF
llama.cpp and compatible local applications
Choose a quantization only after measuring quality and memory.
MLX
Apple Silicon local inference
Measure unified-memory pressure with the intended context.
ONNX
Cross-platform edge and hardware-accelerated deployment
Validate the exported graph and target execution provider.
Continue by implementation decision
Implementation planner
Choose a checkpoint, runtime, device budget, tool boundary, and evaluation gate.
Run locally
Start with Transformers, GGUF, or MLX and verify a reproducible prompt contract.
Base vs. post-trained
Use the official checkpoint names and pick by adaptation versus ready-to-test agent behavior.
Choose a runtime
Map local CPU, Apple Silicon, cross-platform edge, or GPU serving to the published format.
Size the device
Budget weights, KV cache, runtime overhead, tools, and safety margin before claiming fit.
Test tool calling
Validate schemas, permissions, idempotency, traces, and human approval in the target harness.
LFM2.5-2.6B questions, answered
LFM2.5-2.6B is Liquid AI's 2.69B-parameter dense, text-only model post-trained for agentic workloads, native tool calling, and long-context work.
The official model card lists 131,072 tokens. Usable context on a device still depends on runtime, quantization, KV-cache memory, latency, and workload acceptance tests.
Liquid AI names the ready-to-use checkpoint LFM2.5-2.6B and describes it as post-trained for agentic workloads. The separate LFM2.5-2.6B-Base checkpoint is intended for fine-tuning. Use those exact names rather than inventing an Instruct ID.
Use native weights for Transformers or GPU serving, GGUF for llama.cpp-style local inference, MLX for Apple Silicon, and ONNX for cross-platform edge deployment. Confirm the exact repository and runtime version before installation.
Yes. Liquid AI documents native tool calling with tool definitions, a model-produced call, external execution, and a returned tool result. Your application still owns validation, permissions, execution, retries, and approval.
This page does not claim a Flowith integration. Provider weights and documentation do not establish availability or entitlements in a Flowith workspace; check the live model selector.