FrontiersMind has introduced Lumma-Fev-v2, a second generation of open-weight decision models that returns probabilities for typed options instead of generating a sentence for application code to parse. The two announced, available sizes are 4B and 27B, with Hugging Face repositories listing Safetensors under Apache-2.0. Other smaller variants appear in the family roadmap but should not be counted as released merely because they are mentioned in the launch announcement.
A request to Lumma-Fev provides a state, such as text or structured data, together with questions and possible choices. Each possible response receives a probability in one forward pass. That can be useful for support triage, agent routing and human-review queues, where an application needs a consistent contract. It does not make the model's confidence universally accurate or authorize it to perform consequential actions without checks outside the model.
The developers report eight-benchmark mean accuracy of 87.3 percent for the 27B checkpoint and 82.4 percent for the 4B version, against 81.2 percent for TypeSafe Jev 1.13.0 in their comparison. Their report also discusses calibration and individual tasks where the advantage does not hold. These are first-party benchmark claims. Open Model Weights has not rerun the experiment, and a threshold fitted on the publisher's dataset should not automatically be trusted on production requests.
The flagship model is described as a full-weight fine-tune of Qwen3.8-27B with bidirectional full-attention layers while some Gated DeltaNet layers remain causal. The published instructions require a sufficiently recent lumma-fev runtime and an 8,192-token limit for state and questions. The team uses DecisionFormer, which Open Model Weights covered earlier as training infrastructure. Before production deployment, operators should pin the original source, validate calibration on held-out local cases and separate recommendations from actual tool permissions.
What this changes for the evidence layer
Distinct weights from earlier DecisionFormer article. Not a new canonical verified record until separate SHA, weight and license evidence checks.
FrontiersMind — official Lumma-Fev-v2 announcement
This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.
Open source article ↗
