saorsa-1.1-tiny

saorsa-1.1-tiny is Saorsa Labs' low-memory local assistant model for Fae.

• Base model: mlx-community/Qwen3.5-2B-4bit
• Post-training method: ORPO on Fae's assistant, tool-judgment, and memory-preference data
• Intended tier: 8–15 GB Macs
• Intended role: compact local operator with stronger tool use and assistant fit than the base 2B model

Why this model exists

Fae's low-RAM lane needs better tool calling and assistant behavior than a stock compact model can reliably provide. The current saorsa-1.1-tiny retrain improves the base 2B model where the low-memory lane was weakest: tool choice and assistant-fit behavior.

Benchmark delta vs base Qwen3.5-2B

Targeted benchmark gate:

• Tool calling: 9/10 -> 10/10
• Fae capability: 9/20 -> 9/20
• Assistant fit: 7/20 -> 9/20
• Serialization: 9/9 -> 9/9
• /no_think compliance: 5/5 -> 5/5

Artifacts in the Fae repo:

• Base benchmark: scripts/benchmark-results/qwen3.5-2b_targeted_20260314-current.json
• Fine-tuned benchmark: scripts/benchmark-results/qwen35-2b-orpo16fullmlp-exact_targeted_20260314-2004.json

Usage

This model is intended to replace the standard 2B auto-selected lane in Fae while leaving the 4B, 9B, and 27B lanes unchanged.

Notes

• This is a Qwen-compatible MLX model and is intended for local Apple Silicon inference.
• Training data and extraction scripts live in the Fae repository.
• The model is designed for assistant behavior, tool choice, and memory judgment, not generic leaderboard optimization.