ARC-1: a 1.7B decision model answering in ~20ms on a 4060 Ti, free and local
KMatysek · reddit · 2026-10-06
Developer KMatysek spent 11 days training ARC-1, a 1.7B model (Qwen3-1.7B-Base + LoRA) built for typed decisions: support ticket routing, intent detection, moderation, and tool-call gating, outputting choices, scores, or yes/no probabilities.
- Each option is scored in its own branch, so option order doesn't matter; multiple questions about the same text run in one forward pass
- 16ms per short request and 25ms median on JevBench items on a single RTX 4060 Ti at batch 1
- Scores 68.4% on JevBench public 231 (vs. Jev 86.6%, Strands Decider 2B self-reported 72.3%, Laya 58.4%) and 75.5% on DecideBench v1.1
- Weights are CC BY-NC 4.0 (non-commercial training data), code Apache 2.0
The author is upfront that it's clearly less accurate than hosted APIs and publishes all losing numbers; the pitch is speed, local execution, and zero cost. Code, weights, and a Colab quickstart are public.
More from Models
- Mistral launches 1T-parameter MoE model Large 4, claims best open-weights model in US/Europe — Bam4d · 2026-10-06
- Dev's custom vLLM patch runs DiffusionGemma on DGX Spark, edging out the commercial API — vllm_project · 2026-10-06
- Users spot sub-agents in ChatGPT chat mode, hinting at Chat/Work/Codex merger — koltregaskes · 2026-10-06
- Opus 5.5 Builds a Fully Animated Pixel-Art Scene in TypeScript in ~10 Minutes — iamfakhrealam · 2026-10-06
- DINOv3 + SAM Local Vision Pipeline Counts Surgical Instruments in Real Time — MaziyarPanahi · 2026-10-06
- Cohere details Tiny Aya L2-Thinker: data mixing unlocks reasoning in 60 languages — Cohere_Labs · 2026-10-06