October 20, 04:50Pandaily

StartLux-Decision v1.1 model leads Decision Index 0.3.2 ranking among open models

StartLux has released the family of open decision-making models, StartLux-Decision v1.1. The StartLux 27B Instant model now ranks first in the Decision Index 0.3.2 on the Hugging Face platform. StartLux is a company currently in the Genora incubation program.

The Decision Index is a community-supported ranking and is not affiliated with TypeSafe. As of October 16, the 27B Instant model took the top position, achieving a full score of 69.12. It also led in each of the three components of the ranking: 72.88 on public benchmarks, 64.82 in private tests for the same skills, and 61.41 in private tasks from new domains.

The second place was held by the vLLM-SR d3 model with a score of 66.26, and Jev 1.14 achieved a result of 61.53.

These models are designed to perform a narrow practical task. Given a state that can be represented as text, JSON, or images, along with a set of specified questions indicating types, the model outputs a probability for each option: selection from a list, a 'yes' or 'no' judgment, or a level on a scale. Instant models perform one direct pass using bidirectional attention over the query.

AutoThink models work similarly when confident in the answer; however, if the highest probability falls below a set threshold, which defaults to 0.6, they generate a brief chain of reasoning before re-analyzing the answer.

The lineup includes eight open models with parameters ranging from 0.8B to 27B in Instant and AutoThink variants, available in quantized GGUF and MLX formats. They can run on vLLM and MLX. The weights of these models are released under the CC BY-NC 4.0 license, which prohibits commercial use.

On the public subset of Decision Index 0.3, StartLux reports an independently obtained result of 73.37 for the 27B AutoThink model at a threshold of 0.9. Furthermore, the company introduced DecisionPivot—its own benchmark for verifiable decisions, created based on pairs of contrasts that differ by one detail changing the answer, covering policies, registries, graphs, permissions, and tool calls. StartLux claims an accuracy level of 86.9% for the 27B AutoThink model. It should be noted that both these metrics are data provided by the company itself, not independent results.

A technical report published on arXiv under number 2610.17309 describes a three-stage training process and a reinforcement learning method called RLCD-squared. This method rewards more accurately calibrated decision probabilities using correct scoring rules and softly penalizes excessively long reasoning. The authors claim this allowed them to reduce the average length of reasoning by approximately 65% while maintaining high accuracy.

Demonstrations on the project website show how the models perform step-by-step calls in chess, Super Mario Bros., StarCraft II, and in a computer-based shopping task, with a single text prompt given at each step.

Similar stories

nplus1.ruOctober 7, 19:44

OpenAI's internal model provided 372 new solutions in mathematics

OpenAI has posted seven hundred twenty-two mathematical documents on the GitHub platform, which were developed by its internal artificial intelligence model. These manuscripts are grouped into three hundred seventy-two families of results and cover seventeen different areas of mathematics.

Some of the presented proofs were formalized using the Lean language. Among the model's achievements is obtaining a new proof for the Hadwiger–Nelson problem, which is a known problem in combinatorial geometry.

The model demonstrated that it is impossible to color the plane with five colors such that any two points exactly one unit apart have different colors. Consequently, the chromatic number of the plane can now be either six or seven. This proof was also formalized in the Lean system.

pandaily.comOctober 7, 10:06

Shanghai AI Lab releases Intern-Decision: compact models that output decisions instead of text

The Shanghai Artificial Intelligence Laboratory (Shanghai AI Lab) has made Intern-Decision available—a family of compact decision-making models. These models were released on September 28th in three size variants: 0.8B, 2B, and 4B parameters. Instead of generating long textual excerpts, Intern-Decision is designed to directly provide structured judgments along with probabilities for each possible answer.

The design concept is based on the principle of 'decision, not string.' Intern-Decision supports three types of questions. A choice question requires the model to select the best option from a predefined list and assign a probability to each option. A rating question requests a score on a user-defined scale. And a 'yes/no' question returns a binary answer with the corresponding probability. Furthermore, a single query can combine multiple types of questions about the same state, and all are processed in a single forward pass.

Integrating Visual Perception

According to the project repository on GitHub, the models are fine-tuned on the Qwen3.5 language core, while the vision tower and projector remain frozen. This allows Intern-Decision to accept images as part of the input data, meaning visual perception directly influences the decision-making stage. The team demonstrated this capability on tasks related to games and interfaces: the model analyzes raw game frames, including two-dimensional grid maps and side-scrolling levels, and independently selects the next action. Other demonstrations cover mouse control and browser usage.

Testing Results and Performance

The laboratory reports that the 4B model achieves an average accuracy of 90.02% across seven test sets, which is 1.28 percentage points higher than the 88.74% achieved by the commercial decision-making model Jev, which is used as the primary benchmark. These test sets include decision types, tool calling, news classification, and intrusion detection, totaling over 10,000 test strings. The repository also indicates an average local query latency of about 44 milliseconds when using a single consumer RTX 4090 GPU, and provides calibration results on a 96-case distribution benchmark. It should be noted that these figures are self-reported by the developers, and the repository emphasizes that protocols and evaluation scopes differ between models.

The release includes training code, two inference backends, tools for calculating benchmark results, temperature calibration presets, and a browser demo version; however, training data and some closed validation records are not included. Model weights are published on Hugging Face.

Local GPU manufacturer MetaX stated that it has completed the Day-0 adaptation, so Intern-Decision has been running on its hardware since launch. MetaX claims that starting in December 2025, it will provide Day-0 support for 40 major flagship models, and its MXMACA software stack supports over 40 AI frameworks and more than 1000 models.

For developers creating agents, routers, and content filters, small models that output calibrated probabilities instead of free text can make automated decisions more cost-effective, faster, and easier to audit.