ANE Studio Get ANE Studio
THE APPLE SILICON MODEL LABLOCAL / OPEN SOURCE

Small models.
Full control.

Find out what your Mac can do.
Convert small language models to Core ML, verify the result, and run them on your own hardware.

Run from source See the studio

Apple Silicon · macOS 15+ · Source available

FIG. 01 / COMPUTE ASSEMBLY
Conceptual illustration, not measured device activity.
01 / CONVERT02 / VERIFY03 / CHAT04 / COMPARE

THE WORKBENCH

A closer look
at local intelligence.

One place for the model, the conversion, and the evidence. Inspect the work behind every response.

ANE STUDIO / MACOS
ANE Studio benchmark screen comparing CPU, GPU, Neural Engine, and automatic compute policies. Observed placement is marked Not captured.
Actual development screenshot.Example run, not a hardware ranking.

Pinned at the source. Every model snapshot has an exact revision.

Kept on your Mac. Weights, prompts, and results stay local.

Available to inspect. Open-source app. MIT licensed.

FROM CHECKPOINT TO CONVERSATION

Bring a model.
Know what runs.

Conversion is only the beginning. ANE Studio checks the output before you put the model to work.

  1. 01INSPECT + CONVERT

    Start with a checkpoint.

    Inspect Hugging Face metadata, pin a revision, and turn supported weights into a local Core ML package.

    Hugging Facemodel.mlpackage
  2. 02VERIFY

    Check what changed.

    Source and Core ML logits must reach cosine similarity of at least 0.999 and choose the same top token. Failed parity blocks execution.

    cosine ≥ 0.999+top token matches
  3. 03CHAT + COMPARE

    Run the experiment.

    Chat locally and compare CPU, CPU+GPU, CPU+ANE, and automatic policies. Keep raw token timings and planned placement.

    CPUGPUANEAUTO

Current runtime: fixed 64-token context, recomputed each token. No KV cache yet. Architecture ↗

START SMALL

Two checkpoints.
Recorded evidence.

Development conversions tested on an Apple M5 Mac with 16 GB of memory. Each checkpoint earns its own pass.

LIQUID AI / LFM2● PARITY PASSED

LFM2.5 230M

Start here. A small hybrid convolution and attention model.

Logit cosine
0.999936
Top token
Matched
FP16 package
459.8 MB
40cb2ad3View checkpoint ↗
LIQUID AI / LFM2● PARITY PASSED

LFM2.5 350M

A larger checkpoint, tested through chat and policy comparison.

Logit cosine
0.999706
Top token
Matched
FP16 package
709.4 MB
9e6c6ccfView checkpoint ↗

Historical development checks from 29 August 2026. These are conversion results, not a model-quality evaluation or a regression result for later adapters. Model licenses apply separately. Model support ↗

THE IMPORTANT DISTINCTION

A policy is a request.
Evidence is the answer.

Core ML can split a graph across devices or fall back. Selecting the Neural Engine does not prove it ran there.

Read the development results ↗
  1. 01

    Requested

    The devices your compute policy permits.

    Recorded
  2. 02

    Planned

    Operation assignments from MLComputePlan.

    Recorded
  3. 03

    Observed

    An Instruments trace aligned with inference. Profile separately from timing.

    Not captured

MAKE IT YOUR EXPERIMENT

Your Mac.
Your model.
Your studio.

The source is ready to run. There is no public app download linked yet. Start with the commands here, then let ANE Studio prepare your first model.

Apple Silicon · macOS 15+
Bun + uv for source setup. Allow space for model weights and conversion files.

TERMINAL / SOURCE SETUP 
git clone https://github.com/Mr-Neutr0n/ane-studio.git
cd ane-studio
bun install --frozen-lockfile
bun run worker:setup
bun run dev
NEXT ↗

Choose LFM2.5 230M, then Convert to Core ML. The app checks parity before enabling chat.

BEFORE YOU BEGIN

A few details.

Built for experimentation.
Clear about its limits.

Where do my models and prompts go?

Model weights, prompts, converted packages, and results stay on your Mac. Hugging Face receives requests during metadata lookup and downloads. Chat and benchmarks run locally after preparation. Optional credentials use Electron safeStorage and macOS Keychain.

~/Library/Application Support/ANE Studio/

App privacy details ↗

Can I use any Hugging Face model?

No. A checkpoint needs an explicit architecture adapter and a passing conversion. Unsupported architectures fail before weight downloads. The app does not load remote custom model code. The current fixed 64-token context is intended for short experiments.

Is the Mac app signed?

Current builds are unsigned and not notarized. There is no public download linked here yet. Use the source setup, or check the repository for published releases and their installation instructions.

What's open source?

The ANE Studio app is MIT licensed. Downloaded models keep their own licenses and terms. Contributions and checkpoint-level evidence are welcome.

App license · Model licenses