Today
-
v2: Beam reasoning-effort levels (low/medium/high), Accurate variant; v1 is tagged
-
Re-evaluate at a 64k-token limit: resume answers cut off at 32k (corrects AIME comparisons)
2 days ago
-
Real-time M4 Max race video + raw timings; Mac speed note; race and eval scripts
-
Card: link published eval outputs; code: audit_eval.py, export_eval_data.py
-
Card: chart, IFEval, llama.cpp + Space links; code: IFEval, chart and Space builders
-
Dory-Beam-4B weights, config, tokenizer, license and code
