Lumma-Fev-0.1B

Lumma-Fev-0.1B is a decision model. It reads one document (the state) and a set of typed questions, and returns a probability distribution for each question in a single forward pass. It never generates text, so there is nothing to parse and nothing to hallucinate.

It is fine-tuned from FrontiersMind's Nandi-Mini-150M, which we pre-trained from scratch.

NOTE: Lumma-fev-4B and Lumma-feb-9B releasing on 26 September along with a detailed blog.

🎮 Lumma-fev in Action

🔥 See our smallest Lumma-fev 150 million model turn decisions into actions - playing Super Mario autonomously and in real time on a MacBook Air M2. ⚡

⚡ Despite its compact size, Lumma-fev can observe the game state, reason about the situation, and choose the next action — all running locally on consumer hardware. 💻

Benchmarks

Benchmarks

Benchmark / Metric	TypeSafe Jev 1.13.0	Laya	Lumma-Fev-0.15B	Lumma-Fev-0.6B
Banking77	0.87	0.425	0.47	0.90
DAIR Emotion	0.48	0.595	0.68	0.89
AG News	0.91	0.95	0.89	0.85
Typed-decisions	0.72	0.76	0.20	0.30
Average	0.75	0.68	0.56	0.74
P50 Latency	256 ms	32.8 ms	35.576 ms	45.82 ms
Weights	Closed	Open	Open	Open
Cost	$0.042 / 1M tokens	$0 self-hosted	$0 self-hosted	$0 self-hosted

Model details

Developed by	FrontiersMind
Model type	Causal transformer, prefill-only; each question reads the state and its own tokens only; pointer readout over the options
Base model	FrontiersMind/Nandi-Mini-150M (revision 31c8aceaa16e)
Parameters	154M
Fine-tuning	full fine-tune
Pointer head	512 dimensions
Context	up to 1024 state tokens; one question row (state + question) up to 2,048 tokens
Question types	noul (yes/no), choice (1–255 options), score (1–255 ordered levels)
Weights	bf16 backbone, fp32 pointer head
Language	English

Use it with transformers

from transformers import AutoModel

model = AutoModel.from_pretrained("FrontiersMind/lumma-fev-0.1b", trust_remote_code=True)

answers = model.decide(
    state="I was charged twice for my March invoice. Please refund one of them.",
    questions={
        "billing": {"type": "noul", "instructions": "Is this about billing?"},
        "team": {"type": "choice", "instructions": "Which team should handle this?",
                 "criteria": {"billing": "Payments and refunds", "shipping": "Delivery problems", "technical": "Bugs and outages"}},
        "urgency": {"type": "score", "instructions": "How urgent is this?",
                    "criteria": ["can wait", "this week", "today"]},
    },
)
print(answers["team"]["choice"], answers["team"]["probabilities"])

state can be text, a JSON object or an array. Pin a version with revision="<commit>", and move the model to a GPU with model.to("cuda") (bf16 is the stored precision).

Use it with the lumma-fev package

pip install lumma-fev              # local inference
pip install "lumma-fev[serve]"     # plus the API server

import lumma_fev

model = lumma_fev.load("FrontiersMind/lumma-fev-0.1b")        # picks cuda, mps or cpu
print(model.decide("Two charges on my card for one order.",
                   {"billing": {"type": "noul", "instructions": "Is this about billing?"}}))

Serve it as an API

lumma-fev-serve exposes the TypeSafe POST /v1/systemone contract, so existing TypeSafe clients work by changing their base URL.

lumma-fev-serve --model FrontiersMind/lumma-fev-0.1b --host 0.0.0.0 --port 8000
# optional: LUMMA_FEV_API_KEY=<key> requires "Authorization: Bearer <key>"; --cors for browser apps

curl -s localhost:8000/v1/systemone -H 'content-type: application/json' -d '{
  "state": "I was charged twice. Please fix this ASAP.",
  "questions": {"billing": {"type": "noul", "instructions": "Is this ticket about billing?"}}
}'

From Python, with the lumma-fev client or the TypeSafe SDK:

from lumma_fev import Client

client = Client("http://127.0.0.1:8000")
print(client.decide("I was charged twice.", {"billing": {"type": "noul", "instructions": "Is this about billing?"}}))

from typesafe_sdk import Noul, TypeSafeClient     # pip install typesafe-sdk

with TypeSafeClient(api_key="local", base_url="http://127.0.0.1:8000") as ts:
    print(ts.system_one(state="I was charged twice.", questions={"billing": Noul(instructions="Is this about billing?")}).nouls["billing"].noul)

Answers

Type	Criteria	Answer
noul	optional {"true": ..., "false": ...}	noul: probability of yes
choice	{name: description or null}	choice (most likely name), confidence, probabilities by name
score	ordered list of levels	score (expected level), confidence, legend, probabilities by level

Questions never see each other: each one reads the state and its own instructions and options only, so one question cannot change another's answer. Text inside the request cannot forge the model's delimiter tokens.

Limits

• confidence and probabilities are the model's own estimates. Measure calibration on your own labelled data before gating automated actions on them.

Feedback

• Discord: https://discord.gg/ZGdjCdRt
• Email: [email protected]
• Website: https://www.frontiersmind.ai/
• LinkedIn: https://www.linkedin.com/company/frontiersmind/
• X (Twitter): https://x.com/FrontiersMind