Tech News
← Home  ·  All topics

Decision Models

13 GoKawiil briefs on this topic · RSS for decision models

PostHog uses decision models to cut LLM calls in session-recording pipeline

PostHog has deployed small decision models, such as Jev, alongside large language models to handle simpler classification tasks. In its Replay Vision product, the models run two yes/no checks per session summary to detect personal information, letting 98.7% of summaries skip a costlier Gemini LLM call entirely. PostHog says this approach is up to 444.6x cheaper and 193.6x faster than using frontier LLMs directly.

Decision models surge on Vercel's router as Microsoft, startups race to compete

Vercel reports that over a quarter of October requests on its AI model router went to decision models, a new category of LLM that chooses among preset options instead of generating text. TypeSafe's Jev models lead usage with over 1.5 billion weekly queries via OpenRouter, while Microsoft is rolling out its own Microsoft-Decision-1.1 model on Azure, and UK startup Sqwish raised $3 million pre-seed for a decision model with customer-data post-training tools.

Hybrid Jev-Opus model matches frontier accuracy at 4x lower cost

A team tested decision models Jev and Luna against the frontier model Opus on grading identity resolution results, a task that maps scattered data records to real-world identities. Used alone, Jev and Luna scored 88.8-97.3% accuracy, but a hybrid system letting Jev escalate uncertain cases to Opus matched Opus-level accuracy while cutting costs by more than 4x. The tests also found decision models run up to 90 times cheaper and 10 times faster than Opus, though they lag Opus by about 6 percentage points on harder cases.

Open-source spinf-serve lets developers run bulk AI scoring on a single GPU

A developer has released spinf-serve, an Apache 2.0-licensed inference server that reads a piece of content once and answers many fixed-choice questions about it in a single pass, rather than generating text. It runs on a single 24GB NVIDIA GPU, offers an OpenAI-compatible API matching the existing hosted spinf service, and supports both open instruction models and dedicated decision models.

Commentary explores 'make vs buy' economics for AI decision models like Jev 1.13

A blog post examines recent discussion around Jev, a family of AI models accessed via OpenRouter, focusing on its 'decision model' outputs. It describes three output types—choice, score, and noul—and demonstrates them with a sample API call to Jev 1.13 rating review sentiment on a 'Lukewarm' to 'Thrilled' scale.

New 'decision model' AI systems tested on classic Trolley Problem ethics dilemma

A developer ran the Trolley Problem thought experiment through Jev and twelve other 'decision models,' including OpenAI's Luna and eleven other local models, plus a baseline called Laya. These decision models are a new category of AI that output discrete choices with confidence probabilities rather than text, and are marketed as much faster and cheaper than large language models while retaining comparable reasoning ability. The experiment repeated the classic ethical dilemma in multiple phrasings across all models to compare their responses.

ConvAI Innovations founder claims TypeSafe's Jev model copied his 2025 research

After TypeSafe released its Jev decision-model API, which answers yes/no or multiple-choice questions with probabilities instead of generating text, several companies including Microsoft and OpenAI released similar open-source tools. Nandakishor Mukkunnoth of ConvAI Innovations then said he had designed a nearly identical architecture over a year earlier, pointing to his March 2025 arXiv paper on SalesRLAgent, which included published model weights and a dataset.

GoKawiil explains single-pass 'system one' decision models using Qwen3-1.7B

A technical write-up describes how constrained decoding can turn a standard language model into a fast 'decision model' that picks from a fixed set of options (A through E) in a single forward pass, rather than generating output token by token as with structured JSON output. The piece includes example code using Qwen/Qwen3-1.7B to mask the vocabulary down to allowed answer tokens and select the highest-probability one.

Microsoft launches Decision-1 model for fast structured decision tasks

Microsoft has released Microsoft-Decision-1, a model built specifically for decision-scoring tasks like routing, classification, prioritization, verification, and workflow control. It is available now through Microsoft Foundry, with OpenRouter access coming soon. Microsoft reports the model topped a 36-benchmark comparison covering nearly 150,000 questions, running 4.5 times faster than the runner-up, Quyet-1.0-Large, and 35 times faster than GPT-6 Sol.

Jevman benchmark pits AI language models against classic Pac-Man ghosts

Jevman ran 100 games per AI model against scripted Pac-Man ghosts, with each game ending after three lost lives or a 5-minute cap. At every junction the model was asked to weigh the maze, pellets and ghost positions and output movement probabilities, with a 2-second deadline before a backup rule took over. Results were ranked by mean score with a 95% margin of error, and models within that margin were treated as tied.

Strands Labs releases Strands Decider 2B open-source decision model

Strands Labs has launched Strands Decider 2B, a small open-source 'decision model' built for fast local experimentation within its strands-labs initiative. Unlike general-purpose LLMs, it is designed to choose among predefined options or assign numerical scores rather than generate free-form text, and it also outputs a reliability score for each decision. The model is intended to integrate with the Strands Harness SDK and the recently launched Strands harness.

Musubi releases open-weight PolicyLM-1.7B decision model for content moderation

Musubi unveiled PolicyLM-1.7B, a lightweight decision model built to moderate content in under 50 milliseconds using plain-English policies, and released it with open weights. The company says it matches the cost and speed of existing moderation classifiers but can adapt to new rules without retraining, unlike conventional systems.