Lumma-Fev-27B-v2

Lumma-Fev-27B-v2 is a decision model from FrontiersMind, fine-tuned from Qwen3.8-27B. You give it a state (text) and typed questions; it returns a probability for every option in one forward pass. It never generates text, so there is nothing to parse. It works in English, Hindi, Tamil and Kannada (and Hinglish).

We used DecisionFormer for training our model.

What changed from Lumma-Fev-27B: - Bidirectional attention. The full-attention layers read the whole prompt in both directions, like an encoder; the Gated DeltaNet (linear-attention) layers stay causal. - About 300K training samples instead of about 100K.

Results (accuracy, same rows for every model)

None of these test rows appear in the training data. The model was trained on the train splits of AG News, BANKING77, Emotion, RAGTruth and WinoGrande. Financial PhraseBank, JudgeBench and BBH were not in its training data.

Calibration (average over the eight benchmarks, lower is better):

Run it with Ollama

ollama pull FrontiersMind/lumma-fev-27b-v2

Ask a yes/no question about a state from stdin:

ollama run FrontiersMind/lumma-fev-27b-v2 --ask "Does the customer need a reply today?" < ticket.txt

Pick one option, passing each option as name="description":

ollama run FrontiersMind/lumma-fev-27b-v2 \
  --ask "Which team should handle this request?" \
  --choice billing="Charges and refunds" \
  --choice support="Technical integration errors" \
  --choice sales="Questions about buying a product" \
  < ticket.txt

Python

Send a state and a set of named questions to lumma-fev-27b-v2

import ollama

response = ollama.systemone(
    model="FrontiersMind/lumma-fev-27b-v2",
    state="My integration keeps failing. Please help.",
    questions={
        "team": {
            "type": "choice",
            "instructions": "Which team should handle this request?",
            "criteria": {
                "billing": "Charges and refunds",
                "support": "Technical integration errors",
                "sales": "Questions about buying a product",
            },
        },
        "urgent": {"type": "noul", "instructions": "Does the customer need a reply today?"},
    },
)

print(response)

Supported Decision Types

Each question reads the state and its own options only, so one answer never changes another. A question plus the state must fit in 8,192 tokens. The model was trained on noul and choice questions; score questions are accepted but untested.

Hardware

The full-precision weights need about 55 GB of free GPU memory (an 80 GB card). The Ollama build is 17 GB.

Full details are in the model card.

License

Apache 2.0