Nokia’s applied research team has open-sourced AnyJev, a Python library designed to turn an open large language model into a decision model without training. The library focuses on a common production task: choosing one answer from a fixed set instead of writing a sentence.
That difference matters because many AI systems spend a full language-model call on decisions that do not require a written response. AnyJev is built for those choices, while Jev, the decision model created by TypeSafe AI, handles the decision work itself without generating text.
A simpler path from open LLMs to decisions
AnyJev installs from PyPI and ships under the Apache-2.0 license. It includes transformers and vLLM backends, along with shared-prefix scoring, giving developers a way to use the library with open LLM systems already supported by those backends.
The library turns an open LLM into a decision model without requiring a new training process. Its target is practical: select one result from a known set, such as a routing choice or classification outcome, instead of producing an open-ended answer.
Jev takes that idea into a dedicated model. TypeSafe AI released Jev on September 15, 2026, describing it as a “System One” decision model. The company built Jev to take unstructured state as input and return structured, probabilistic decisions in a single parallel pass.
Jev does not generate text. Each decision includes a probability score, so an enterprise can set thresholds for automatic execution, human review, or escalation. Its outputs are schema-guaranteed, with no JSON parsing failures or malformed responses.
Speed, cost, and structured reliability
TypeSafe’s published benchmarks state that Jev runs at 70–500 milliseconds end to end, compared with 3–329 seconds for frontier LLMs. Jev costs $0.042 per million input tokens, while output is free.
TypeSafe also states that Jev has a 0% structured output error rate, compared with up to 45.5% for some frontier models. For comparable LLM decision tasks, Jev is listed as up to 445 times cheaper, reducing decision costs to fractions of a cent and latency to less than 500 milliseconds.
The benchmark figures point to a specific use case rather than a replacement for every language-model task. Jev is aimed at fast, repeated judgments with a defined output structure, while general-purpose LLMs remain responsible for tasks that need extended reasoning, text generation, or open-ended dialogue.
TypeSafe’s figures also state that auto-decidable traffic at a 5% error level rises from 7.7% to 52.0%. In practical terms, that means more decisions can move through an automatic path when the system can attach a confidence score and apply a clear threshold.
GPTBots.ai splits decisions from reasoning
Aurora Mobile Limited announced on September 22, 2026, that its enterprise AI agent platform, GPTBots.ai, had integrated Jev. The integration arrived within days of Jev’s September 15 release.
GPTBots.ai now divides its architecture into two parts. A Decision Layer powered by Jev handles fast, high-volume judgments in under 500 milliseconds and at a fraction of the cost of an LLM call. A Reasoning Layer uses general-purpose LLMs for complex reasoning, text generation, and open-ended dialogue.
The platform’s existing Model Auto-Router, Dynamic Top-K, and Intent Classification features in FlowAgent and Workflow are now powered by Jev. These functions can make repeated choices before a larger model handles the part of the task that needs more explanation or generation.
Chris Lo, Founder and CEO of GPTBots.ai, described the change this way: “What Jev gives us is a dedicated decision engine that handles these judgments at a speed and cost point that changes the economics of the entire pipeline. Instead of paying for a full LLM call on every routing decision, we now pay a fraction of a cent per judgment — and get calibrated confidence scores to boot.”
Why the combination matters
Jev has also been integrated by Vercel, Cloudflare, LangChain, and other major developer platforms. Alongside AnyJev’s open-source library, those integrations place decision-focused AI in both developer tooling and enterprise agent systems.
Aurora Mobile, founded in 2011, is a global provider of customer engagement and marketing technology services. Its platform includes Omnichannel Engagement, AI-Driven Marketing, Advanced AI Customer Support, and Frictionless Identity Security. Its flagship brand is EngageLab, while GPTBots.ai forms part of its AI infrastructure, and the company trades on NASDAQ as JG.
The broader idea is straightforward: not every AI step needs a long answer. A model that can make a structured choice, show its confidence, and return it within 500 milliseconds can take care of the pipeline’s routine decisions, leaving general-purpose LLMs to handle the work that truly needs language and reasoning.
Based on
