---
title: "Invergent's Open-Weight Rune v3 Hits Second on the Decision Index"
subtitle: "Invergent's open-weight Rune v3 pulls neck-and-neck with Jev 1.13 on the Decision Index and takes the top spot on image decisions."
url: "https://alphasignal.ai/news/invergent-s-open-weight-rune-v3-hits-second-on-the-decision-index"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "AlphaSignal"
company_url: "https://alphasignal.ai/company/alphasignal"
published: "2026-09-21T17:58:18.000Z"
updated: "2026-09-30T20:02:08.870Z"
type: "Model"
topics:
  - "Llms"
  - "Open Source"
subtopics:
  - "Mixture Of Experts"
  - "Vision Language"
  - "Small Models"
upvotes: 107
source_url: "https://huggingface.co/surogate/rune-26b-a4b-GGUF"
comments: 0
truncated: true
---

# Invergent's Open-Weight Rune v3 Hits Second on the Decision Index

> Invergent's open-weight Rune v3 pulls neck-and-neck with Jev 1.13 on the Decision Index and takes the top spot on image decisions.

# Rune v3 ranks second on the Decision Index

Invergent’s open-weight [Rune v3 weights](https://huggingface.co/surogate/rune-26b-a4b-GGUF) score 57.44 on the public Decision Index, placing second behind the proprietary Jev 1.13 at 57.89\. An estimated result with Rune’s optional thinking mode reaches 59.2, although that figure sits outside the standard leaderboard. The Apache 2.0 checkpoint turns text, structured data, or images into a selected option and a probability distribution, making it suitable for routing, classification, tool selection, and other constrained decisions.

## One call, three decision shapes

Rune accepts a state, a question, and a set of permitted answers. For each question, the default path selects one answer and assigns probabilities to every option in a single forward pass. Its context window extends to 262k tokens, and the API supports three question types:

* **`choice`:** Select one of several options.
* **`noul`:** Estimate the probability that a statement is true.
* **`score`:** Select a position on an ordinal scale.

The checkpoint uses `Gemma4ForConditionalGeneration` with a mixture-of-experts architecture. It contains 26.5 billion parameters and routes each token through 8 of 128 experts. The weights are distributed as bf16 safetensors; despite the repository name, v3 GGUF quantizations are not yet available.

## A 0.45-point race at the top

The [Decision Index](https://huggingface.co/spaces/multimodalart/jev-decision-index) v0.2 suite contains 151,034 requests drawn from 44 benchmarks. Its headline score uses a 38-benchmark panel, rescales each benchmark so random-chance accuracy equals zero, averages results within five areas, and weights those areas into one index. Higher scores are better.

__Headline and selected area scores__
| System                                | Index | Knowledge & Reasoning | Tools & Automation | Arts & Human Taste |
| ------------------------------------- | ----- | --------------------- | ------------------ | ------------------ |
| Rune v3, thinking enabled (estimated) | 59.2  | 48.0                  | 72.4               | 41.3               |
| Jev 1.13                              | 57.89 | 51.3                  | 75.1               | 37.7               |
| Rune v3, bf16                         | 57.44 | 43.4                  | 71.2               | 41.9               |
| AutoJev-27B                           | 56.40 | 40.9                  | 79.3               | 39.4               |
| simple-jev, Qwen3.8-27B               | 55.74 | 36.6                  | 76.2               | 36.5               |

The standard bf16 checkpoint trails Jev 1.13 by 0.45 points. No uncertainty interval accompanies these results, so the ranking alone does not establish statistical significance. Rune trails Jev in Knowledge & Reasoning and Tools & Automation, while leading it in Retrieval & Classification and Arts & Human Taste. The estimated thinking result would place Rune 1.31 points ahead of Jev.

_Preview only. Full text requires an AlphaSignal Pro account._

## Sources

- [surogate/rune-26b-a4b-GGUF on Hugging Face](https://huggingface.co/surogate/rune-26b-a4b-GGUF)
- [surogate engine on GitHub](https://github.com/invergent-ai/surogate)
- [Jev Decision Index leaderboard](https://huggingface.co/spaces/multimodalart/jev-decision-index)