AGILLM 603B (bin8) inference weights

Backup of the AGILLM binary MoE run (agillm128b-moe-local-v1, 603,471,216,640 binary weights) at training step 6234, as an exact 1-bit sign pack.

Everything is under packs/step000006234/. The format, layout and usage are in its README.md.
model_bits.bin and table_bits.bin are the weights (75.84 GB), ckpt.pt holds the small floating-point state, and trainer_source.py is the single-file trainer and inference program.
SHA256SUMS lists the SHA-256 of the pack files.

The pack is inference-exact. It does not contain the native int8 latents, so resuming training from it is a warm restart.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support