Lumma-Fev-v2: Open Decision Models, Benchmarks and Local Deployment

FrontiersMind announced Lumma-Fev-v2 on October 29, releasing 4B and 27B decision models that return probabilities over defined answers rather than generate text. For developers routing requests or classifying documents, the useful change is a model designed around the decision itself, without a paragraph to parse afterward.

The official announcement describes support for English, Hindi, Tamil, Kannada and Hinglish. These are downloadable models, not a new general-purpose chatbot.

What changed in Lumma-Fev-v2

The main architectural change is bidirectional full attention. Those layers can read the entire prompt in both directions, while the models’ linear-attention layers remain causal. That distinction matters: calling the whole architecture bidirectional would overstate what changed.

For the 27B model, FrontiersMind also reports expanding the training set from roughly 100,000 to 300,000 examples. The interface still takes a state, such as a support request, and typed questions with defined answer options.

Consider a billing queue. An application could ask whether a request concerns a payment and which team should handle it. It receives probabilities and a selected option, rather than asking a language model to write JSON and then checking whether the generated response follows the schema.

A constrained answer format removes one integration problem, not the possibility of a wrong decision. The application still needs thresholds, evaluation data and a route for uncertain cases.

Local deployment requirements

The 4B model card documents the smaller checkpoint’s requirements. The 27B card specifies a considerably larger memory budget.

Model	Documented BF16 GPU memory	Input limit	Required package
Lumma-Fev-4B-v2	About 10GB	8,192 tokens	lumma-fev 0.3 or newer
Lumma-Fev-27B-v2	About 55GB	8,192 tokens	lumma-fev 0.3 or newer

The input limit includes the question and state. Both models have Apache 2.0 licensing, but downloadable weights do not make inference free: the hardware, serving and operational work remain yours.

The version requirement is functional, not cosmetic. Older package versions do not apply the required attention behavior. Developers loading weights manually should follow the model-card instructions rather than assume that changing a configuration flag alone is enough.

How to use the API

The lumma-fev package supports local loading and a self-hosted, TypeSafe-compatible HTTP endpoint at POST /v1/systemone. That is a deployment option, not evidence of a separately priced hosted API.

The package’s October 30 releases also document support for a smaller 0.1B-v2 model, although the original announcement still describes that model as forthcoming. For the 4B and 27B models, the documented trained question types are yes/no decisions and choices between options. Ordered score questions are accepted but remain untested, so they should not quietly become production assumptions.

What the benchmarks do and do not establish

The 27B model card reports improved accuracy and calibration, but discloses an important qualification: training includes the training splits of five of the eight benchmark datasets.

That does not establish test-set leakage. It does mean those results mix familiar task distributions with held-out tasks, so the headline average should not be read as a clean measure of unfamiliar-workload performance. The published comparisons are also the developer’s evaluations, not independent validation.

For a routing system, a useful next test is narrower: does the model send fewer requests to the wrong queue at the same human-review rate? Better calibration only becomes operationally valuable when the probabilities help make that trade-off.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn model releases and product updates into clear stories developers can understand and share.

Book a 30-minute call.