voxor-ai/gemma-4-e2b-it-coreml

Gemma 4 E2B instruct, Core ML

google/gemma-4-E2B-it as compiled Core ML models for the Neural Engine (gemma-4-E2B-it-coreml/chunk1..4.mlmodelc, run in that order; int8 weights, fp16 compute, iOS 18 / macOS 15 or later), with int8 embedding tables, RoPE tables and the tokenizer.

• Context 4096 tokens, sliding window 512; hidden_states [1, 1, 1536], per-layer input [1, 1, 8960]
• chunk1 and chunk2: functions decode_q1 (1 token) and verify_qK (256 tokens); chunk4 ends in the argmax token id
• Embeddings: [262144, 1536] and [262144, 8960] int8 per row with fp16 scales
• Files over 1 MB are zstd-compressed (<file>.zst); decompress before use

Accuracy: on 11 prompts (27 to 1500 words, English and German), greedy next-token agreement with a Q4_0 quantization of the same model is 98.3 % (2568 / 2612 tokens).

Apache License 2.0; see LICENSE and NOTICE.