Introduction

TensorFlow does not have a hard limit of five models. What people usually mean is that TensorFlow's checkpoint retention defaults to five saved checkpoints in a rolling window. In TensorFlow 1 this is controlled by Saver(max_to_keep=5), and in TensorFlow 2 the same idea is handled by CheckpointManager.

What the Default of Five Actually Means

In TensorFlow 1 style code, tf.compat.v1.train.Saver keeps the most recent checkpoints according to max_to_keep. The default is five, which is why older checkpoints disappear during training unless you change that value.

1import tensorflow as tf
2
3
4tf.compat.v1.disable_eager_execution()
5
6w = tf.compat.v1.get_variable("w", initializer=0.0)
7update = tf.compat.v1.assign_add(w, 1.0)
8saver = tf.compat.v1.train.Saver(max_to_keep=5)
9
10with tf.compat.v1.Session() as sess:
11    sess.run(tf.compat.v1.global_variables_initializer())
12    for step in range(10):
13        sess.run(update)
14        path = saver.save(sess, "./ckpt/model", global_step=step)
15        print(path)

After enough saves, only the newest checkpoint files remain in the active retained set. That is a retention policy, not a model-count ceiling.

Change the Retention Policy Intentionally

If you need more recovery points, raise max_to_keep.

saver = tf.compat.v1.train.Saver(max_to_keep=20)

If you need fewer, reduce it. The right number depends on how often you save, how expensive training is, and how much storage you can afford.

Some TensorFlow 1 documentation also notes that None or 0 changes deletion behavior, but that still does not create a five-model hard limit. It just changes how checkpoint bookkeeping works.

The TensorFlow 2 Approach

Modern TensorFlow code should use tf.train.Checkpoint together with tf.train.CheckpointManager.

1import tensorflow as tf
2
3model = tf.keras.Sequential([tf.keras.layers.Dense(1)])
4optimizer = tf.keras.optimizers.Adam(1e-3)
5
6ckpt = tf.train.Checkpoint(step=tf.Variable(1), optimizer=optimizer, model=model)
7manager = tf.train.CheckpointManager(ckpt, directory="./tf2_ckpt", max_to_keep=5)
8
9for _ in range(8):
10    ckpt.step.assign_add(1)
11    path = manager.save()
12    print("saved", path)
13
14print("latest", manager.latest_checkpoint)

CheckpointManager makes the retention policy explicit. According to the TensorFlow API docs, it keeps some checkpoints and deletes unneeded ones, oldest first, until only max_to_keep remain in the active set.

Keep Best and Rolling Checkpoints Separate

A rolling retention count is good for crash recovery, but it is not the same thing as preserving the best model by validation metric. If training quality fluctuates, the best-performing checkpoint may be older than the newest five.

A practical setup is:

• rolling checkpoints for recent recovery
• a separate "best model" checkpoint when validation improves
• optional long-term milestone snapshots for audits or reproducibility

1best_loss = float("inf")
2current_val_loss = 0.12
3
4if current_val_loss < best_loss:
5    best_loss = current_val_loss
6    ckpt.write("./best/model")

This prevents a good model from being deleted simply because later checkpoints were saved afterward.

Storage and Restore Planning

Checkpoint retention is an operational decision, not just a code parameter. Large models can consume significant disk space. If you increase retention aggressively, also plan for storage monitoring and cleanup policies.

More importantly, test restoration regularly.

1latest = manager.latest_checkpoint
2if latest:
3    ckpt.restore(latest)
4    print("restored", latest)

A checkpoint strategy is only useful if restore actually works during recovery.

Common Pitfalls

The biggest misunderstanding is treating the default value of five as a hard TensorFlow limit. It is only the default retention count.

Another issue is increasing max_to_keep without considering storage growth. More checkpoints improve recovery history, but they also cost disk space.

Teams also sometimes rely only on rolling checkpoints and forget to preserve the best-performing model separately.

Finally, be careful when mixing TensorFlow 1 and TensorFlow 2 checkpoint styles in the same project. The retention idea is similar, but the APIs and object models are not the same.

Summary

• TensorFlow does not have a hard limit of five models.
• In TensorFlow 1, Saver defaults to max_to_keep=5.
• In TensorFlow 2, CheckpointManager provides the same retention concept.
• Retention count should match recovery needs and storage budget.
• Keep best-model checkpoints separately from the rolling checkpoint window.