LFM2.5-2.6B Base vs. Instruct: Choose the Right Checkpoint

Quick answer

The official choice is Base versus post-trained, even though searchers may say “Base vs. Instruct.” Liquid AI currently publishes:

Official checkpoint	Provider description	Start here when
LiquidAI/LFM2.5-2.6B-Base	Pre-trained base model for fine-tuning	You own a custom training and evaluation program
LiquidAI/LFM2.5-2.6B	Post-trained for agentic workloads	You want to test chat, reasoning, and native tool calling

There is no need to invent LFM2.5-2.6B-Instruct as a repository ID. Use the exact provider name in code, manifests, evaluations, and deployment records.

Choose the post-trained checkpoint first when

• the task is tool use, extraction, RAG, or a long-context workflow;
• you need a baseline before deciding whether custom training is justified;
• the prompt, retrieval, schema, or tool layer may solve the gap;
• you do not yet have rights-cleared training data and an immutable holdout;
• you need to validate runtime and device fit before training work.

The post-trained checkpoint is still not a production approval. Liquid AI does not recommend it for agentic coding or knowledge-heavy tasks, and your test must include failures and permission boundaries.

Choose Base only with a training contract

Before selecting Base, write down:

1. the behavior the post-trained checkpoint failed to meet;
2. the legal and product basis for each training asset;
3. train, validation, and untouched holdout splits;
4. the adaptation method and versioned hyperparameters;
5. base-versus-custom quality, regression, safety, latency, and memory gates;
6. artifact lineage, access controls, removal path, monitoring, and rollback;
7. the target runtime and compatible export format.

A lower training loss does not establish a better agent. Evaluate completed tasks, correct tool selection and arguments, unsupported claims, refusals, recovery, and final-answer grounding.

Run a fair comparison

Keep the same task set, context, tool schemas, device class, and acceptance metrics. Treat custom training as one intervention alongside prompt, retrieval, tool description, and workflow changes. Approve the smallest change that reliably meets the requirement.

Use the local setup guide for a baseline and the tool-harness guide for agent evaluation.

Frequently asked questions

Is there an official LFM2.5-2.6B-Instruct checkpoint?

Liquid AI’s current model card names LFM2.5-2.6B-Base and LFM2.5-2.6B. It describes the latter as post-trained for agentic workloads; do not invent an Instruct repository ID.

When should I use LFM2.5-2.6B-Base?

Use Base when you intend to fine-tune a custom checkpoint and have rights-cleared data, an untouched evaluation set, compute and operations capacity, and a deployment approval process.

When should I use LFM2.5-2.6B?

Use the post-trained checkpoint when you want to evaluate Liquid AI’s ready-to-test agent, tool-calling, extraction, RAG, and long-context behavior before considering custom training.

Official sources

• LiquidAI LFM2.5-2.6B model card
• LiquidAI LFM2.5-2.6B-Base repository
• Liquid LFM2.5-2.6B documentation

Source check: August 5, 2026. Verify current checkpoint names, revisions, licenses, training support, and runtime compatibility before use.