AGILLM-8B (bintf2, training in progress)

Binary (1-bit packed) 8B language model, 7.92B parameters, trained with bintf2 (int8-latent sign training) on FineWeb-Edu.

• Tokenizer: DeepSeek, vocab 129,280
• Exports: ~1.1 GB packed one-bit weights, roughly every 500 steps / 60 min
• Layout: exports/agillm8b_bits_step[_].pt + .receipt.json (sha256, provenance)
• Trainer code: trainers/

Training is in progress; early checkpoints are not useful models yet.