Kev
A family of lightweight decision models built on Qwen open-source models, focused on structured judgment tasks such as routing decisions, intent classification, and content filtering. The models are positioned as low-cost, fast-response, small-scale specialized models.
Timeline (last 90 days)
Direct use of Kev achieves 64.00% accuracy, and Qwen3.5-4B outputting only the final answer achieves 68.80% accuracy.
Kev is a judgment-model-based reasoning framework/interface that allows invoking a frozen model to score candidate answers via natural language or structured instructions, without relying on the model to directly generate a complete answer.
The ABACUS team instantiated the framework with Kev: the teacher model plans reasoning steps and generates program code, and the frozen judgment model acts as a semantic executor that scores candidate answers for each step.
Kev is an open-weight decision model family associated with developer Jared Palmer, with the same task positioning as Jev: answering predefined questions about the input and outputting structured results.
Kev's documentation supports fine-tuning on your own labeled samples, and its server can use a request format compatible with Jev, so migration cost is relatively low.
Commit #18976 enables MLX to load the Kev model, and to simultaneously load its FP32 decision head and release metadata.
Commit #18976 supports System One scoring on top of the shared state cache, and adds Kev-specific resource limits.
Ollama released the v0.40.4-rc1 pre-release, with the core change being the introduction of Kev decision scoring capability in the MLX backend.
Open-source projects attempting to replicate Jev's capabilities include nano Jeff, a project formerly known as open Jeff, and a family of micro decision models named Kev.
Laya, Kev, and Ollaya are three open-source alternatives frequently mentioned in the context of Jev.
8 more timeline events
All Facts (15)
Kev is a judgment-model-based reasoning framework/interface that allows invoking a frozen model to score candidate answers via natural language or structured instructions, without relying on the model to directly generate a complete answer.
50%UnverifiedDirect use of Kev achieves 64.00% accuracy, and Qwen3.5-4B outputting only the final answer achieves 68.80% accuracy.
50%UnverifiedThe ABACUS team instantiated the framework with Kev: the teacher model plans reasoning steps and generates program code, and the frozen judgment model acts as a semantic executor that scores candidate answers for each step.
50%UnverifiedKev is an open-weight decision model family associated with developer Jared Palmer, with the same task positioning as Jev: answering predefined questions about the input and outputting structured results.
50%UnverifiedKev's documentation supports fine-tuning on your own labeled samples, and its server can use a request format compatible with Jev, so migration cost is relatively low.
50%UnverifiedOpen-source projects attempting to replicate Jev's capabilities include nano Jeff, a project formerly known as open Jeff, and a family of micro decision models named Kev.
50%UnverifiedLaya, Kev, and Ollaya are three open-source alternatives frequently mentioned in the context of Jev.
50%UnverifiedFor the policy-intensive customer service routing pilot task, Kev is a stronger starting point, while Laya is faster and has lower memory usage but requires improvement under these rules before it can be trusted.
50%UnverifiedKev is an open-source classification model based on Qwen 3.5's 0.8B base model, with LoRA fine-tuning weights merged in, using unidirectional reading and making decisions through a pointer head.
50%UnverifiedTested on the same machine equipped with a Spark CPU and limited to three threads, Laya's median call latency was 541 milliseconds and Kev's was 756 milliseconds, with the 95th percentile at approximately 607 and 802 milliseconds respectively.
50%UnverifiedUnder the original option order in the customer service ticket routing test, Kev scored 26/30 and Laya scored 12/30.
50%UnverifiedAcross three option orderings combined, Kev scored 74/90 and Laya scored 33/90; Kev had 4 messages and Laya had 5 messages where results changed with the option order.
50%UnverifiedIn cases requiring contextual understanding, Kev correctly handled the situation where a customer explicitly stated they did not want a refund and only needed a password reset by selecting technical support, while Laya incorrectly selected billing due to the presence of billing-related vocabulary.
50%UnverifiedKev is a family of lightweight decision models built on Qwen3.5, positioned as "Jev-like" decision models.
50%UnverifiedKev is an open-source reproduction project built on Qwen 3.5, offering decision models in three sizes: 0.8B, 4B, and 9B.
50%