Sam

Local model I specify and train. Dual RTX 3090 at home. Newest daily is the one that runs. Sam 3.0 is not shipped.

• Sam 2.7 thinkon Current. Qwen3.8 27B class. GRPO merge on the 2.6 line plus DAPO-GRPO on the gap pack. Think stays on. Daily local model, 2026-09.
• Sam 2.6 GRPO line that 2.7 merged. Gap-pack training, not a public daily.
• Sam 2.5 Full-epoch QLoRA on official Qwen3.8-27B. Eval loss 0.656. Think blocks kept. Smoke 55.8 tok/s. Became the daily coder, 2026-08-16.
• Sam 2.0 Telegram front door. Personal agent on the Mac, 2026-07-31. Not a chat-app clone.
• Sam 1.5 Continuum QLoRA from 1.2. Thinking on and off retest. Q8 ship on llama.cpp, 2026-07-30.
• Sam 1.2 First PC ship after cloud QLoRA. Hard-suite refuse beat V1. About 127 tok/s Q8 on the dual 3090 path, 2026-07-29.

Tovar Works LLC · Anchorage