The model's developer released DecisionTune 1.0, an Apache-2.0 395M encoder (ModernBERT-large plus a 4 KB scoring head) that scores given options rather than generating text, aimed at routine agent decisions like tool routing and yes/no checks. It runs fully locally via PyTorch, MLX, or ONNX backends and ships with an MCP server for local assistants. The developer is seeking feedback on where it picks wrong.
ARC-1 is a small decision model that takes context plus a question and returns a choice, score, or yes/no probability for tasks like ticket routing, intent detection, moderation, and agent tool calls. It is built by fine-tuning Qwen3-1.7B-Base with LoRA and runs in ~20 ms on a single consumer GPU. It targets developers building agentic or classification pipelines that need fast, cheap structured decisions.
ARC-1 is a small decision model that takes context plus a question and returns a choice, score, or yes/no probability for tasks like ticket routing, intent detection, moderation, and agent tool calls. It is built by fine-tuning Qwen3-1.7B-Base with LoRA and runs in ~20 ms on a single consumer GPU. It targets developers building agentic or classification pipelines that need fast, cheap structured decisions.
A developer released Decido-1, a family of open models that answer yes/no and multiple-choice decisions inside agents, returning probabilities per option instead of generated text. The 26B MoE (4B active) scores 85.4 on the six benchmarks used for Jev 1.13 (API model, 83.9) and runs on a laptop CPU in about 4 seconds per 1k-token doc with 5 questions. Weights are on Hugging Face, installable via pip with GGUFs for llama.cpp.
A developer released Decido-1, a family of open models that answer yes/no and multiple-choice decisions inside agents, returning probabilities per option instead of generated text. The 26B MoE (4B active) scores 85.4 on the six benchmarks used for Jev 1.13 (API model, 83.9) and runs on a laptop CPU in about 4 seconds per 1k-token doc with 5 questions. Weights are on Hugging Face, installable via pip with GGUFs for llama.cpp.
A developer released mlx-decision, a free MIT-licensed Python package that runs open decision models (Cloudflare Clef, Perplexity pplx-decider, d3, H2O-Lightning, LiquidAI d1) locally on Apple Silicon via MLX. It supports fine-tuning, model comparison, and dataset evaluation through a web Studio, Python, CLI, HTTP API, and MCP. Ported models run 1.4–6x faster than the original PyTorch code on MPS.
A developer released mlx-decision, a free MIT-licensed Python package that runs open decision models (Cloudflare Clef, Perplexity pplx-decider, d3, H2O-Lightning, LiquidAI d1) locally on Apple Silicon via MLX. It supports fine-tuning, model comparison, and dataset evaluation through a web Studio, Python, CLI, HTTP API, and MCP. Ported models run 1.4–6x faster than the original PyTorch code on MPS.
A developer released Sabaki, an MIT-licensed router that turns unstructured input into typed decisions (choice/score/noul) with calibrated confidence, running entirely on CPU. It uses a TF-IDF keyword router over tiny specialists with a cross-encoder fallback, trading accuracy for millisecond latency and offline use. The author shared benchmark numbers and calibration failures openly and is soliciting community critique.
A developer released Sabaki, an MIT-licensed router that turns unstructured input into typed decisions (choice/score/noul) with calibrated confidence, running entirely on CPU. It uses a TF-IDF keyword router over tiny specialists with a cross-encoder fallback, trading accuracy for millisecond latency and offline use. The author shared benchmark numbers and calibration failures openly and is soliciting community critique.
A developer released Mica v0.1 4B, an Apache-2.0 decision model for agent gates and routers that returns calibrated probabilities instead of generating text. Built on Qwen3.5-4B with a rank-16 LoRA and trained on 77,732 rows for under $30 of rented RTX 3090 time. It scores below the closed Jev 1.13 API on benchmarks but resists in-state prompt injection far better (69% vs 18%) and runs in about 47 ms per decision on an RTX 3090.
A developer released Mica v0.1 4B, an Apache-2.0 decision model for agent gates and routers that returns calibrated probabilities instead of generating text. Built on Qwen3.5-4B with a rank-16 LoRA and trained on 77,732 rows for under $30 of rented RTX 3090 time. It scores below the closed Jev 1.13 API on benchmarks but resists in-state prompt injection far better (69% vs 18%) and runs in about 47 ms per decision on an RTX 3090.