kirp/jpt-35b-a3b

TEXT GENERATIONPricing:Input $0.4 / Cached $0.07 / Output $4Concurrent Unit Cost:2Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 27, 2026License:cc-by-nc-4.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

kirp/jpt-35b-a3b is a 35.1 billion parameter open decision model developed by kirp, built upon Qwen/Qwen3.5-35B-A3B. It specializes in providing calibrated probabilities for options in response to typed questions (choice, score, noul) from a given situation, with only 3 billion parameters active per token. This model excels in general typed decision-making, achieving 0.892 on JevBench v1.4 public and 52.89 on Decision Index 0.2.1, making it suitable for applications requiring fast, probabilistic decision outputs without generative explanations.

Loading preview...

JPT-35B-A3B: An Open Decision Model

JPT-35B-A3B is the largest in the JPT series, developed by kirp and built on the Qwen3.5-35B-A3B mixture-of-experts architecture. This model is designed to take a situation and typed questions, returning calibrated probabilities for every option in a single forward pass. Unlike traditional LLMs, it focuses purely on decision outputs without generating explanations or reasoning tokens, optimizing for latency.

Key Capabilities

  • Typed Decision Interface: Implements the Jev typed-decision interface, supporting choice (2-255 labels), score (ordered scale), and noul (yes/no) question types.
  • Efficient Processing: Only 3 billion of its 35 billion parameters are active per token, contributing to its speed.
  • Vision Integration: Retains the vision tower from its base model, allowing it to process screenshots, photos, and video frames alongside text.
  • Strong Benchmarking: Achieves a public accuracy of 0.892 on JevBench v1.4.0 and a score of 52.89 on Decision Index 0.2.1, outperforming its base model and other alternatives in typed decision tasks.

Good For

  • Applications requiring rapid, probabilistic decision-making from structured inputs.
  • Scenarios where a calibrated confidence score for each option is more valuable than a generated explanation.
  • Integrating with systems like llm2jev for streamlined decision processing.
  • Use cases involving visual context, as its vision capabilities are preserved.

Limitations

  • Weakest in arithmetic and date-related tasks, as it lacks an explicit reasoning phase.
  • Can be overconfident in-distribution, with a typed-decisions test ECE of 0.171.
  • Primarily English-focused, with other languages having limited coverage.
  • Requires significant hardware (66 GB bf16 weights) for hosting, necessitating multiple GPUs.

Finetunes
1 models