Lumma-Fev-27B
FrontiersMind's 27B Apache-2.0 decision model that scores supplied options in one forward pass, reporting 90.7% vs TypeSafe Jev's 83.9%.
Overview
Lumma-Fev-27B, announced October 14, 2026, is the largest model in FrontiersMind's Lumma-Fev family of decision models, which score options you supply in a single forward pass instead of generating text to parse. It is built on Qwen3.8-27B, trained on about 100K decision samples in English, Hindi, Tamil and Kannada, and handles yes/no (probability of yes) and choice (one label from up to 255 options) questions. On 4,454 held-out decisions FrontiersMind reports 90.7% accuracy versus 83.9% for TypeSafe Jev and calibration error of 0.009 versus 0.045, while Jev remains ahead on reasoning benchmarks such as BBH and WinoGrande. Weights are about 49 GiB in bfloat16 and fit on a single 80 GB GPU.
This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.
Key features
- Single-forward-pass scoring for yes/no and up-to-255-option choice questions
- 90.7% accuracy vs 83.9% for TypeSafe Jev on 4,454 held-out decisions (FrontiersMind-reported)
- Calibration error 0.009 vs 0.045 for Jev
- Trained on ~100K samples in English, Hindi, Tamil and Kannada; 3.2 hours on four H200 GPUs
- Apache 2.0; ~49 GiB in bfloat16 on one 80 GB GPU
- pip install lumma-fev; HTTP server via lumma-fev-serve
Use cases
- Routing and classification decisions inside agent workflows
- Multilingual (Indic) decision-making where calibrated probabilities matter
Pricing
Free, open weights under Apache 2.0.
Pricing reflects the source record and may change. Confirm current terms with the provider.
Dates and sources
- Released
- 14 Oct 2026
- Record updated
- 19 Oct 2026
- Record last checked
- 19 Oct 2026