danielamitay/gemma-4-e2b-it-lut8-tensor-swev

gemma-4-e2b-it-lut8-tensor · Swev

A Core ML export of google/gemma-4-E2B-it for Swev, a Swift package for local typed decisions. Returns choices, ordinal scores, and probabilities for runtime-defined questions; this export does not generate free-form text.

• Package: gemma-4-e2b-it-lut8-tensor-swev-l4096-k16.mlpackage (download the entire directory).
• Precision: 8-bit per-tensor weight palettes with FP32 computation.
• Text context: up to 4,096 tokens including the question, options, and formatting. Swev selects the smallest fitting bucket: 128, 256, 512, 1,024, 2,048, or 4,096.
• Questions: choice, score, and noul; up to 16 answer options per question and 64 questions per request, evaluated independently.
• Runtime: Swift 6, macOS 15+ or iOS 18+, with a Swev version supporting schema 2.0. Export version 0.3.0. Validated locally on macOS with CPU-only Core ML; memory and compute-unit support depend on the device.

Text and single-image inputs share one package. Text-only requests skip the vision graph. The image route has a separate 256-token capacity, including 64 image tokens, and uses one PNG/JPEG resized to 384×384 with aspect-fit white padding. Audio, video, multiple images, and the source model’s full context window are not included.

Palettization compresses supported weight tensors; some tensors remain FP32. This is LUT weight compression, not NVIDIA NVFP4/NVFP8 arithmetic.

Swift usage

Add Swev as a Swift package dependency, then:

import Swev

let model = try await SwevModel.load(from: HuggingFaceModel(
    repository: "danielamitay/gemma-4-e2b-it-lut8-tensor-swev",
    package: "gemma-4-e2b-it-lut8-tensor-swev-l4096-k16.mlpackage"
))
let response = try await model.predict(
    state: "apple",
    questions: [.noul(id: "edible", instructions: "Is this an edible food?")]
)
print(try response.noul("edible").noul)

The first call downloads the package; later calls reuse Swev’s local download cache. Keep the loaded model resident to avoid recompilation and initialization on each request. See loading and caching and model support. No Python runtime or separate tokenizer/adapter files are required.

Source and conversion

Source checkpoint: google/gemma-4-E2B-it at 3e22461f65e8. Tokenizer, formatting, and inference settings are bundled in the package. This is a converted export, not a new fine-tune.

The export keeps text and vision paths, restricts output to candidate-label logits, and omits audio. It preserves the 512-token sliding attention window in the corresponding text layers.

Conversion tooling and the package contract are documented in Swev’s conversion guide and schema reference. Exported capacity is not a guarantee of task accuracy; evaluate the checkpoint on your own inputs.

License

Apache License 2.0; see LICENSE. Source model attribution and pinned revisions are listed above. This is an independent Swev-compatible export, not an official release by the original model authors.