manjunathshiva/opendecider-nano

Zero-Shot Classification • 0.4B • Updated about 13 hours ago • 279 • 3

Note Start here · ~400M, 17 ms per question on a GPU, also fast on CPU · typed-decisions 0.796 (Laya-td 0.766, both fine-tuned) · pip install opendecider

Note Try it in your browser, no install

manjunathshiva/opendecider-small

Zero-Shot Classification • Updated about 13 hours ago • 110

Note 4B LoRA on Qwen3-4B for decisions it has never seen: 0.735 zero-shot on general decisions (Jev 0.730), ECE 0.087 · fits a 16 GB Mac

manjunathshiva/opendecider-small-td

Zero-Shot Classification • Updated about 13 hours ago • 62 • 1

Note 4B tuned for business workflows (support, invoices, security alerts, agent traces) · typed-decisions 0.792

manjunathshiva/opendecider-medium-td

Zero-Shot Classification • Updated about 13 hours ago • 32

Note 30B MoE, 3B active · most accurate self-hostable model on general decisions (0.765) · typed-decisions 0.788 · NVIDIA, multi-GPU

manjunathshiva/opendecider-large-td

Zero-Shot Classification • Updated about 13 hours ago • 15

Note 80B MoE, 3B active · best calibration (ECE 0.083) · typed-decisions 0.801 · NVIDIA, multi-GPU

manjunathshiva/opendecider-small-GGUF

Zero-Shot Classification • 4B • Updated about 13 hours ago • 117

Note opendecider-small for LM Studio, Ollama and llama.cpp · Q8_0 gives the full-precision top answer on 98.8% of typed-decisions · Q4_K_M for small machines

manjunathshiva/opendecider-small-td-GGUF

Zero-Shot Classification • 4B • Updated about 13 hours ago • 128 • 1

Note opendecider-small-td for LM Studio, Ollama and llama.cpp · Q8_0 agrees with full precision on 98.6%

manjunathshiva/opendecider-small-mlx-8bit

Zero-Shot Classification • 4B • Updated about 13 hours ago • 45

Note Apple Silicon (MLX) · 4.5 GB · same answer as full precision on 1,955 of 2,000 · 66 ms per question

manjunathshiva/opendecider-small-mlx-4bit

Zero-Shot Classification • 4B • Updated about 13 hours ago • 31

Note Apple Silicon with little memory · 2.6 GB · typed-decisions 0.651

Note The typed-decisions benchmark: nano and the -td models were fine-tuned on its train split only; the test split was never seen