Lumma-Fev-27B-v2

Lumma-Fev-27B-v2 is a decision model. 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). It is a full-weight fine-tune of Qwen3.8-27B and the successor of Lumma-Fev-27B.

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)

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

Lumma-Fev-27B-v2 was trained on the train splits of AG News, BANKING77, Emotion, RAGTruth and WinoGrande, so those five are in-distribution for it. Financial PhraseBank, JudgeBench and BBH are not in its training data.

Use it

Use the lumma-fev package, version 0.3 or later: this checkpoint needs bidirectional attention, which older versions do not apply (they refuse to load it). You need a GPU with about 55 GB free (an 80 GB card), or several smaller GPUs with device_map="auto".

pip install "lumma-fev>=0.3"

import lumma_fev

model = lumma_fev.load("FrontiersMind/Lumma-fev-27B-v2")       # or lumma_fev.load("27b-v2")
# model = lumma_fev.load("27b-v2", device_map="auto")          # several GPUs: pip install "lumma-fev[large]"

answers = model.decide(
    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(answers["team"])
# {"type": "choice", "choice": "support", "confidence": 0.99..., "probabilities": {"billing": ..., "support": ..., "sales": ...}}

state can be text, a JSON object or an array. Each question reads the state and its own options only, so one answer never changes another. Images: pass images=[...] (paths, PIL images or data: URLs) after pip install "lumma-fev[images]". 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.

To serve it behind the TypeSafe-compatible POST /v1/systemone API:

pip install "lumma-fev[serve]"
lumma-fev-serve --model 27b-v2 --host 0.0.0.0 --port 8000       # add --device-map auto for several GPUs

If you load the weights yourself: decision_config.json has "attention_mode": "noncausal_full_attention". Set config.is_causal = False on the text config and is_causal = False on every text attention module after loading; setting only the config makes single unpadded requests run causally.

For speed, also install the flash-linear-attention and causal-conv1d kernels; without them transformers uses slower PyTorch code.