News
Key events in decision models since Jev launched the category, in order, each with first-hand sources.
StartLux-Decision v1.1 Released: Pioneering AutoThink Architecture and RLCD², Leads Decision Index by a Landslide with 73.4 Points
StartLux officially releases the open-source decision model family StartLux-Decision v1.1 (spanning 0.8B–27B plus MoE variants, arXiv:2610.17309). It pioneers a dual-form design of Instant (single-forward-pass fast intuition) and AutoThink (confidence-adaptive short thinking), and introduces the RLCD² reinforcement learning algorithm that jointly penalizes and calibrates probabilities and thinking length. It scores 73.4 on the Decision Index 0.3 public leaderboard, holding a landslide lead of nearly 10 points over similar models; it simultaneously open-sources the adversarial evaluation benchmark DecisionPivot and the full set of GGUF/vLLM deployment weights.
Other major events this month
10-15Liquid AI Releases d1-ultra: The Strongest Multimodal Decision Model, Leading Across All Domains10-14Microsoft Releases First Decision Model Microsoft-Decision-1 as Big Tech Accelerates Entry10-01Cloudflare Open-Sources Clef / Clef-flash and Launches RL Fine-Tuning PlatformOctober 2026
11 eventsStartLux-Decision v1.1 Released: Pioneering AutoThink Architecture and RLCD², Leads Decision Index by a Landslide with 73.4 Points
StartLux officially releases the open-source decision model family StartLux-Decision v1.1 (spanning 0.8B–27B plus MoE variants, arXiv:2610.17309). It pioneers a dual-form design of Instant (single-forward-pass fast intuition) and AutoThink (confidence-adaptive short thinking), and introduces the RLCD² reinforcement learning algorithm that jointly penalizes and calibrates probabilities and thinking length. It scores 73.4 on the Decision Index 0.3 public leaderboard, holding a landslide lead of nearly 10 points over similar models; it simultaneously open-sources the adversarial evaluation benchmark DecisionPivot and the full set of GGUF/vLLM deployment weights.
Liquid AI Releases d1-ultra: The Strongest Multimodal Decision Model, Leading Across All Domains
Liquid AI's flagship multimodal decision model d1-ultra follows d1. Built by scaling up a large open-source foundation, it natively processes text and images in a single parallel forward pass (e.g., comparing product spec sheets against defects, UI visual review), with input pricing unchanged at $0.042/million tokens (output free). It surpasses Jev 1.14 and d1 across the Decision Index and calibration accuracy (ECE), topping multiple leaderboards.
Microsoft Releases First Decision Model Microsoft-Decision-1 as Big Tech Accelerates Entry
Microsoft has officially launched Microsoft-Decision-1 on Microsoft Foundry and OpenRouter. Built on Qwen3.5-9B with targeted reinforcement fine-tuning, it specializes in agent termination and branching decisions, LLM routing, and enterprise-grade safety guardrail classification. Input costs $0.042 per million tokens; it outputs calibrated probabilities in a single forward pass without generating long-form text.
S1MB Benchmark Suite Fully Upgraded and Restructured: Consolidated to 164 Benchmarks, Evaluated Models Expanded to 179
S1MB maintainer hotchpotch optimized the benchmark organization, consolidating 58 scattered Japanese NLP evaluation sets into 27 standardized core tasks and a Generalization Japanese suite, bringing the full suite to 164 benchmarks; evaluated models surged to 179, with evaluation records surpassing 23,600.
TypeSafe AI Upgrades Jev 1.14 and Launches Proprietary Jev Router
Category pioneer TypeSafe AI releases Jev 1.14, comprehensively strengthening long-context input and multimodal experimental judgments, while also launching Jev Router, optimized for code generation and chat routing, helping developers quickly complete routing decisions from complex intents to the best downstream LLM in a single forward pass (within tens of milliseconds).
OpenRouter's Decision Catalog Surpasses 20 Models: Respan Span-02 and Inception Mercury Decide Listed
OpenRouter's decisions output-modality catalog continues to expand, newly listing open-weight Respan Span-02 (specialized for agent monitoring and output evaluation) and Inception Mercury Decide (high-frequency process control), showing a trend toward ecosystem diversification and specialization.
Decision Index 0.3 Released: Perplexity Decider v1.1 Takes the Top Spot, Jev Drops to 3rd
The new version weights are 0.2×public + 0.5×same_skills + 0.3×new_domains. Perplexity Decider v1.1 (27B) ranks first with 62.8 points, Fastino GLiDE second with 60.2, and Jev 1.13 third with 60.1. The same release also introduced a Vision sub-leaderboard, where JEV-27B-VL leads with 69.8 points.
Intern-Decision 0.8B/2B/4B Released: Fine-tuned Language Backbone, Frozen Vision Tower
An open-source multimodal decision model from the Shanghai AI Lab lineage; one <decision> token per field, roughly 33 ms/query on a single RTX 4090, with training code open-sourced as well.
Cloudflare Open-Sources Clef / Clef-flash and Launches RL Fine-Tuning Platform
Clef 27B (Qwen3.8-27B base) and Clef-flash 9B (Qwen3.5-9B), Apache-2.0, with a vision encoder and 64K context; scores all options in parallel in a single forward pass. Training uses label-smoothed CE + Brier loss for calibration, plus RLCD. This is the first decision model with both open weights and an official fine-tuning service.
Perplexity Decisions API Documentation Released, Directly Reusing S1MB's noul/choice/score Naming
Endpoint POST /v1/decisions, model pplx-decider-v1.1-27b, supports images (base64), $0.02 per million input tokens with free output — the cheapest among all decision endpoints. Its v1 weights have been open-sourced under Apache-2.0.
OpenAI launches Decisions API (public beta), multimodal too
Endpoint POST /v1/decisions, only one model gpt-6-luna, supports text + image input, three query modes: predicate / choice / score. Input $0.10 per million tokens, output free. Supports ZDR and HIPAA, regions limited to US and Europe.
September 2026
9 eventsbekko-system-one-v0 Released: 17M / 68M / 400M, Runs in the Browser
An extremely small decision model based on the Ettin-reranker (ModernBERT); the 17M ONNX+INT8 version is only about 29 MB, and completes 26,269 judgments in just 13 seconds on an RTX 5090. The 400M version outperforms many 10B+ models on S1MB. Training code, 153 training subsets, and hyperparameters are all open-sourced.
S1MB (System One Mosaic Benchmark) Released: 137 Benchmarks, 106 Data Subsets
Built by bekko author hotchpotch, focusing on three decision types — Noul / Choice / Score — covering 14,009 cases and 26,269 judgments, plus 6 additional synthetic generalization benchmarks. Evaluation code, datasets, and the adapter protocol are all open-sourced.
Liquid AI releases d1: surpasses Jev on the Decision Index for the first time
d1 scores 58.9 on Decision Index 0.2.1, while Jev 1.13 scores 57.9. Compatible with the three primitives, leading in artistic judgment, language understanding, and retrieval classification; slightly lower on tool calling, and 8 points behind on knowledge reasoning.
Together AI releases Tev1-4B-experimental and opens its training recipe
LoRA SFT on Qwen3.5-4B (rank 8, 1 epoch, lr 5e-5), 37,840 train / 4,568 val; the company says you can train your own for about $17. Note: the weights license is still not finalized.
Jev Decision Index Launches: 132,422 Requests, 37 Benchmarks
A third-party-maintained leaderboard for decision models. House rules forbid prompt tuning, truncation, and dropping options; unanswered or errored items are scored 0 followed by chance correction. This is the first comparable horizontal evaluation for the category.
Local runtimes explode: laya-mlx, laya-coreml, laya.cpp, jevmlx
The community is bringing decision models to Apple Silicon and edge devices, with short decisions in 5–14 ms on M3 Max. The "no text generation" design makes inference extremely lightweight.
Laya (Convai Innovations) and Kev (Jared Palmer) Released
Two of the most important open-source decision model families. Laya uses ModernBERT-large, focusing on edge deployment and millisecond-level latency; Kev spans 0.5B–27B with the best-documented fine-tuning path (0.8B needs only 4 GB of VRAM).
Community Starts Replicating Within Days: Mapika decider, pngwn typed-decisions, OpenDecision and More
Within 24 hours of Jev's release, the first open-source replications and typed-decision datasets appeared on Hugging Face. To date, a total of 101 original checkpoints / 67 model families have accumulated.
TypeSafe AI Releases Jev 1.13, Proposes System One Decision Models
The first decision model that returns typed probabilities without generating text. Its three primitives Choice / Noul / Score have become the de facto standard. Input costs $0.042 per million tokens, output is free, and the official claim is latency of 70–500 ms.
Mirrored daily from decision.host · translations and summaries are AI-generated; the original is linked on every item.