Noema
Noema
Get Noema

Models

Using Noema Overfit

Run compatible mixture-of-experts models using a resident core and locally paged expert weights, then measure whether the model is practical on your device.

Verified for
Noema 3.5+
Applies to
iPhone / iPad / Mac / Vision Pro
Last reviewed
July 20, 2026

How Overfit works

Noema Overfit lets compatible mixture-of-experts models run even when their complete weights cannot fit in device memory.

Overfit is not model training or “overfitting.” It reorganizes a compatible GGUF model into a complete .noema-paged package containing:

  • resident.gguf, with the weights required throughout inference.
  • One or more experts-*.bin files containing routed expert weights.
  • manifest.json, describing the package structure, model geometry, file hashes, expert records, and integrity information.

During inference, Noema keeps the resident core and a device-sized expert bank in unified memory. When the model selects an expert that is not already present, Noema reads it from local storage, verifies it, and loads it into the bank.

Download and run an Overfit model

  1. Open Explore and select the GGUF catalog.
  2. Open the Noema Overfit paged-model collection.
  3. Select a complete paged package and review its storage size, resident memory, expert-bank estimate, context limit, and estimated working set.
  4. Download the package. Noema treats the manifest, resident GGUF, and expert files as one model.
  5. Open the model in Stored, then open its settings.
  6. Leave Overfit (Paged Experts) set to Automatic.
  7. Run the Canary Test to measure the model on this device.
  8. Load the model and begin a chat.

Create a package on Mac

Noema for macOS can convert an existing compatible MoE GGUF into a complete paged package.

  1. Download and register the source GGUF.
  2. In Stored, right-click the model.
  3. Choose Create Paged Package…
  4. Wait while Noema extracts and verifies the expert data.
  5. Choose Add to Stored when conversion finishes.

The result is a sibling <model-name>.noema-paged directory. Conversion does not retrain, prune, or replace the model’s weights.

The Canary Test

The Canary Test performs a short, repeatable local run for this package and device.

  • Validates a sample of the package.
  • Measures cached and uncached storage reads.
  • Loads the paged runtime.
  • Runs a fixed 64-token completion.
  • Records time to first token and generation speed.
  • Measures latency percentiles and long stalls.
  • Records expert-bank hit rate, misses per token, peak memory, and thermal state.

The result belongs to the exact package fingerprint, device, storage volume, native runtime contract, and app build. Available memory is evaluated separately at launch.

Understand the Canary result

  • Paged — interactive
  • Paged — slow
  • Too slow on this device
  • Relay recommended
  • Not supported for this model

Current compatible models

As of July 20, 2026, the public Noema Overfit repository contains these complete packages.

Model packageApproximate downloadResident coreExpert payloads
Gemma 4 26B-A4B QAT, UD-Q4_K_XL14.36 GB1.40 GB12.96 GB
Qwen 3.6 35B-A3B, UD-Q4_K_M22.15 GB2.57 GB19.57 GB
Qwen 3.5 122B-A10B, Q4_K_M74.27 GB4.00 GB70.26 GB

The runtime whitelist currently covers qwen3moe, qwen35moe, and gemma4. This includes the currently published Qwen 3.5, Qwen 3.6, and Gemma 4 packages. Dense models and arbitrary GGUF architectures are not supported.

Privacy and safety

After the model is downloaded, Overfit inference can remain entirely local.

  • Model weights remain on local storage.
  • Prompts and responses remain on the device.
  • No cloud inference provider is required.
  • Expert payloads are excluded from device backups.
  • Matching conversation-prefix states can be saved locally for faster follow-up launches.
  • Relay remains a separate, optional route.

Packages are checked for valid geometry, safe paths, file sizes, record bounds, coverage, fingerprints, and checksums. The native runtime validates the package independently at launch.

Under memory pressure, Noema can reduce prefetching, discard optional queued work, or stop generation before an unsafe memory condition.

Current limitations

  • Overfit is restricted to compatible MoE architectures.
  • Package creation is currently macOS-only.
  • Paged launches currently cap context at 8,192 tokens on Mac and 4,096 tokens on iPhone, iPad, and Vision Pro.
  • Multimodal projectors and some utility workflows do not use the paged path yet.
  • Low Power Mode disables expert prefetching to favor energy use.
  • Prompt-state restoration only works when the package and relevant runtime configuration still match.
  • Force (Experimental) may bypass performance recommendations, but it never bypasses package-integrity validation.