SNE

SNE is a sealed local-inference engine for Apple silicon. It runs an admitted Gemma 4 12B serving contract through Apple MLX, MLX-C, and Metal without a Python runtime in the serving path.

Read the launch article: SNE v1: A 59.185 tok/s Local Gemma 4 12B Breakthrough on Apple M5

This repository is public for distribution, verification, issue tracking, and reproducibility. The SNE engine is closed-source software distributed as a sealed binary under the SNE Technical Preview License.

SNE v1 Technical Preview

The first public package is qualified for this exact contract:

• Apple M5, arm64
• macOS 26.5.2
• AC-power test session
• mlx-community/gemma-4-12B-it-8bit
• pinned target snapshot 200bb6db075e137a4deb08838865ac4ddb86292e
• google/gemma-4-12B-it-assistant
• MTP-assisted decode with exact continuation and terminal-logit gates

The immutable admission campaign completed 100 of 100 exact runs. Its median generation rate was 59.185 tok/s (minimum 58.291 tok/s, final ten-run median 59.152 tok/s). These are Technical Preview measurements on the qualified machine and contract, not a guarantee for other hardware, models, contexts, or power states.

The reported 752.61 GB/s value is an effective checkpoint-byte proxy. It is not a measured physical DRAM, cache, or fabric-bandwidth counter.

Install

Download sne-v1.0.0-technical-preview-darwin-arm64.tar.gz and its checksum from the v1.0.0 Technical Preview release.

shasum -a 256 -c sne-v1.0.0-technical-preview-darwin-arm64.tar.gz.sha256
tar -xzf sne-v1.0.0-technical-preview-darwin-arm64.tar.gz
cd sne-v1.0.0-technical-preview-darwin-arm64
./tools/verify.zsh

This Technical Preview is ad-hoc signed and is not Apple-notarized. Do not disable macOS security controls to run it. A Developer ID-signed and notarized distribution is a later release gate.

Run the qualified contract

SNE does not distribute model weights. Obtain them under their respective licenses, then provide local paths:

export SIRSI_MODEL_PATH=/absolute/path/to/gemma-4-12B-it-8bit-snapshot
export SIRSI_MTP_ASSISTANT_PATH=/absolute/path/to/model.safetensors

./tools/doctor.zsh
./bin/sne --prime-only
./bin/sne

--prime-only must pass before a scored run. SNE fails closed if its admitted continuation gate does not match.

What this release is and is not

• It is a local, sealed, binary-only inference Technical Preview.
• It is MTP-assisted serving; 59.185 tok/s is not a plain non-speculative affine-8 decode claim.
• It contains no model weights, telemetry, call-home service, DRM, or Python serving runtime.
• It does not claim measured GPU occupancy, DRAM traffic, SLC traffic, MMU activity, register pressure, or ANE activity where public counters do not expose those quantities.
• It is not an open-source release.

See claims and nonclaims, reproducibility, and architecture boundary.

Background

SNE grew from the M5 Silicon Observatory, a correctness-first investigation of Apple-silicon inference surfaces. The project first established exact-token native parity and a stable 40+ tok/s non-speculative serving lane, then admitted MTP only after exact continuation and terminal-logit checks. The v1 package is the first immutable, reproducible product boundary from that work.

Apple, MLX, Metal, Gemma, and Google are trademarks of their respective owners. SNE is an independent Sirsi product and is not affiliated with or endorsed by Apple or Google.