Krummholz drops Atlantic v8 open weights

Calibrating open-weight models to explicitly say "I don't know" is a refreshing departure from confident confabulation, though trading away harmless completions for strict safety remains a delicate balance.

• –Self-play alignment: The weights were trained exclusively on model-generated outputs ranked and filtered by code, bypassing human annotation and synthetic teacher LLMs.
• –Measurable hallucination reduction: On 229 zero-tool factual prompts, Atlantic Lite 4B dropped wrong answers from 52% in v7 to 2% in v8.1, opting for "I don't know" on 141 questions.
• –High-reliability tool calling: Both models achieved 100% valid tool formatting (60/60) and answered zero questions incorrectly from memory when web search was enabled.
• –Quantization penalty: Aggressive 4-bit (Q4_K_M) mobile quantization bumped zero-tool errors from 2% to 10%, indicating that truthfulness calibration degrades under heavy compression.
• –Open Apache 2.0 licensing: Developers can freely self-host and fine-tune both checkpoints ahead of the planned January 2027 stable release.