Alibaba Introduces ThinkJev to Cut AI Reasoning Costs for Typed Decisions

Alibaba researchers introduced ThinkJev, an open framework built on Qwen3.5 open hybrid-thinking models that calibrates reasoning compute for typed decisions. By dynamically choosing between fast direct execution (System 1) and full chain-of-thought reasoning (System 2), ThinkJev slashes latency and token usage while maintaining target accuracy standards across enterprise decision workflows.

What Is Alibaba's ThinkJev Framework for Qwen3.5?

In modern generative AI deployments, many practical tasks resolve into a typed decision—a choice among a predetermined set of discrete options, such as multi-choice classifications, mathematical verifications, dates comparisons, or categorical routing. Traditional large language models (LLMs) can extract probabilities for these choices in a single forward pass, but direct answers frequently fail on complex, multi-step logic.

While enabling extended reasoning—commonly referred to as chain-of-thought or reasoning tokens—significantly increases accuracy, it incurs substantial latency and inference costs. Alibaba's research team addressed this trade-off with ThinkJev, a dual-system inference framework applied to Qwen3.5 models (including 4B, 9B, and 27B parameter variants). ThinkJev leverages an open hybrid-thinking architecture where the same prompt can yield either an immediate direct response or a calibrated, step-by-step reasoning chain without repeating prompt prefill or losing probability alignment.

How Does ThinkJev Balance Speed and Accuracy Across System 1 and System 2?

ThinkJev structures decision-making into two distinct operational modes derived from cognitive psychology concepts:

• System 1 (Fast Execution): Reads option probabilities directly from the model's initial forward pass. This takes milliseconds and uses zero additional output tokens.
• System 2 (Slow Reasoning): Reads probabilities after generating a thought process from a cached prompt, ensuring seamless escalation without repeating token processing.

What sets ThinkJev apart is its unified probability space. Because option tokens (e.g., A, B, C, D) are evaluated identically whether sampled immediately or after a reasoning step, developers can sample a thought process and evaluate the optimal answer under any operational budget. Without requiring additional task fine-tuning, ThinkJev applies statistical calibration on dev sets to set tight error bounds. For instance, at a strict 5% target error rate, Qwen3.5-27B resolves 24.5% of decisions instantly via System 1 while keeping real-world error down to 3.9%.

What Were the Performance Benchmarks Achieved by ThinkJev on JevBench?

For complex queries requiring deeper reasoning, ThinkJev dynamically monitors confidence after each thought token and halts generation as soon as certainty exceeds a statistically sound threshold. On testing datasets, this adaptive stopping strategy reached 81.6% accuracy (compared to 82.7% when always forcing full reasoning) while requiring 72% fewer thought tokens.

Furthermore, on public items of JevBench—an independent benchmark dedicated to evaluating typed-decision AI services—Qwen3.5-27B equipped with ThinkJev triggered deep reasoning in only 29.4% of evaluation cases. Despite using a fraction of the compute, its accuracy proved statistically indistinguishable from commercial proprietary thinking APIs.

Key Operational Takeaways for Enterprise AI Pipelines

• Adaptive Compute Allocation: Compute budgets should reflect query complexity rather than uniform max-token limits.
• Statistically Guaranteed Accuracy: Calibration algorithms allow enterprises to maintain target error rates (e.g., under 5%) deterministically.
• Prompt Caching Efficiency: Reusing cached prompts between instant choices and deep reasoning eliminates redundant compute overhead.
• Open-Source Parity: Open models like Qwen3.5 combined with smart inference frameworks match closed proprietary APIs at a fraction of the cost.