The Pulse
FrontiersMind Releases 27B Model for Fixed-Choice Decisions
FrontiersMind released Lumma-Fev-27B, an open-weight model that returns choices and probabilities for user-defined decisions. The company reports 90.7% accuracy on 4,454 held-out decisions, while noting that TypeSafe Jev performed better on
AI.info Team · Written with AI ·
FrontiersMind released Lumma-Fev-27B on October 14, a model built to choose among answers an application defines rather than generate a label in prose. A developer supplies a state—such as a support ticket—and one or more typed questions; the model returns a choice and probabilities in a single forward pass. The release targets fixed-outcome work such as routing requests or applying yes-or-no gates, where software can use structured results without parsing generated text.
One model, caller-defined options
Lumma-Fev-27B supports English, Hindi, Tamil and Kannada. Its `choice` mode selects from a caller’s labels, while `noul` returns a probability for a yes-or-no decision. Each question reads the shared state and its own options independently, according to the model card. The repository lists Apache 2.0 licensing and a maximum of 8,192 tokens for the state and question together.
FrontiersMind says the model starts from Qwen3.8-27B and received a full-weight fine-tune, with about 100,000 decision samples. The team trained it in a single pass over the data. The release supports `score` questions in its interface, but the authors say the model was not trained on that type; they advise treating those results as untested.
90.7% on the company’s decision test
On 4,454 held-out decisions, FrontiersMind reports 90.7% accuracy for Lumma-Fev-27B, compared with 83.9% for TypeSafe Jev on the same rows. The company also reports an expected calibration error of 0.009, against 0.045 for Jev, on that test set. Calibration measures whether stated confidence corresponds to observed accuracy; the reported result applies to the company’s test distribution, not automatically to a different product or user population.
“On reasoning benchmarks such as BBH and WinoGrande, Jev is still ahead.”
Abhay Kumar, Vishesh Tripathi, Ramsha and Rohit, FrontiersMind
The authors compare the models across eight open benchmarks as well as the held-out decision set, and say Jev leads on reasoning tasks including BBH and WinoGrande. Their results therefore describe a specialized advantage on the tested choice-based decisions, not a general lead across model capabilities. FrontiersMind also says there is no public benchmark for decision-making in Indic languages; its Hindi, Tamil and Kannada results come from an internal benchmark.
The confidence score still needs local testing
A probability output can let a team route high-confidence cases automatically and send uncertain ones for review, but the release does not establish safe thresholds for any particular workflow. FrontiersMind tells users to fit thresholds using labeled examples from their own traffic before gating actions on scores. That distinction matters when a wrong routine classification and an incorrect approval carry different costs.
The weights are available through Hugging Face, but running the full bfloat16 model requires substantial memory: the company lists about 49 GiB of weights and recommends an 80 GB GPU for a single-device setup. Its model repository says the checkpoint is not deployed through an inference provider. Teams testing it locally will need to supply the hardware and validate performance on their own workloads.