--- license: other library_name: jax tags: - jax - orbax - test-time-training - early-exit --- # E2E 1B p24-r1: full_network_binary_router This is a weights-only, inference-ready Orbax release of the terminal step-1024 E2E checkpoint. - Hidden size: 2048 - Layers: 24 - Training sequence length: 8192 - Router z coefficient: 0.0001 - GTCF: true - Evaluation: NIAH-1: complete, NIAH-2: complete, NIAH-3: complete, LongBenchV2: complete, PG-19: complete `PG-19: complete` is an explicit status, not a numeric score. Use `checkpoint/` as the manager root and restore step `1024` through the project's read-only routed inference loader. The package includes only `model_weights`; it cannot resume optimizer training. No tokenizer bytes, optimizer state, data-iterator state, tracking metadata, internal service paths, or raw evaluation samples are included. Use a separately obtained compatible Llama-3-base tokenizer. See `LICENSE_NOTICE.md` before use or redistribution.