PurpleMIST-Mini-2.0

Flash-level decisions at half the size, now reading images.

PurpleMIST-Mini-2.0 is the newest model in our flagship decision line. Give it a message, ticket, record, receipt or screenshot and any number of typed questions about it, and it returns a calibrated probability for every possible answer in a single pass, in 19 languages.

🤗 Get the weights

Previous generation:PurpleMIST-Flash-1.0 PurpleMIST-Mini-1.0

0.723

LocalLLaMA/typed-decisions, English zero-shot

46.5

Jev Decision Index, public text suite

63.7

Jev Decision Index, public vision suite

4.5B

Parameters

/ Why it exists

The problem it solves

Decisions do not only arrive as text. A refund claim comes with a photo of the receipt, a bug report with a screenshot, a payment dispute with a picture of the transfer. Until now that meant one model to read the image and another to decide.

Mini-2.0 reads the image and the text together and answers every typed question about them in one pass, with calibrated probabilities. It matches PurpleMIST-Flash-1.0 on text at about half the size, so the same decisions cost less to run.

/ What you get

Highlights

Flash accuracy at half the size

0.723 on LocalLLaMA/typed-decisions (English, zero-shot) against 0.720 for the 7.9B Flash-1.0, and 46.5 on the Jev Decision Index public suite against Flash’s 46.7.

Reads images

63.7 on the Decision Index public vision suite, 6th of 24 models up to 6B parameters. 0.984 accuracy on held-out receipts and 0.907 on screenshots and real photos. A state can carry several images.

19 languages, asked across languages

English, eight African languages and ten other major languages. Questions can be in a different language from the state, at 0.937 accuracy on held-out cross-language questions.

Long inputs, read whole

Trained on states up to 98,304 tokens and never truncates: on the Decision Index all 140,178 text requests were answered whole, at a median 70 ms each.

/ Quick start

Run it in a few lines

$pip install "transformers>=5.22" "torch>=2.10" huggingface_hub pillow

decide.py

import os, sys
from huggingface_hub import hf_hub_download

repo = "olaverse/PurpleMIST-Mini-2.0"
sys.path.insert(0, os.path.dirname(hf_hub_download(repo, "purplemist.py")))
from purplemist import Decider

d = Decider.from_pretrained(repo, device="cuda")

state = {"channel": "whatsapp",
         "message": "Ẹ jọ̀ọ́, wọ́n ti yọ owó lẹ́ẹ̀mejì lórí káàdì mi fún ọjà kan náà. Mo fẹ́ kí ẹ dá owó mi padà."}
questions = {
    "team":    {"type": "choice", "instructions": "Which team should handle this?",
                "criteria": {"billing": "payments, charges and refunds",
                             "delivery": "orders in transit", "technical": "app or account problems"}},
    "refund":  {"type": "noul", "instructions": "The customer is asking for money back."},
    "urgency": {"type": "score", "instructions": "How urgent is this?",
                "criteria": ["can wait", "normal queue", "today", "immediately"]},
}
d.decide(state, questions)

/ PurpleMIST family

Which PurpleMIST model?

Coming to the Olaverse API as its decision model. Need every last point on reasoning-heavy questions? PurpleMIST-Flash-1.0 (7.9B) is still a little ahead there.

	PurpleMIST-Mini-2.0THIS PAGE	PurpleMIST-Flash-1.0 →	PurpleMIST-Mini-1.0 →
Parameters	4.5B	7.9B	1.9B
LocalLLaMA/typed-decisions (English, zero-shot)	0.723	0.720	0.601
African Typed Decisions (8 languages)	0.636	0.668	0.416
Jev Decision Index, public text suite	46.5	46.7	13.1
Images	Yes	No	No
Languages	19	English and 8 African	English
Best for	Text and images at lower cost	Reasoning-heavy and African-language text	English on small hardware
Generation	Current	Previous	Previous
On the Olaverse API	Coming soon	No, self-host only	No, self-host only

/ Where it shines

Best for

Checking receipts, invoices and screenshots: does the total add up, what kind of purchase is it, is this a genuine payment confirmation.
Routing and triaging tickets, chats and emails in English, African languages and other major languages.
Moderation, compliance and KYC checks where you need a probability, not just a label.
Long inputs such as transcripts, logs and case files, read whole in one pass.

/ Built with

Qwen3.5-4B-Base Transformers PyTorch Apache-2.0

More production models

All models

MIST-Extract

PurpleMIST-Mini-2.0

The problem it solves

Highlights

Flash accuracy at half the size

Reads images

19 languages, asked across languages

Long inputs, read whole

Run it in a few lines

Which PurpleMIST model?

Best for

More production models

MIST-Extract-1.0

MIST-Extract-Flash-1.0

lid-neural-608

lid-lite-608

diacnet-2.0

diacnet-mini-2.0

diactag-2.0

mist-qg-1.5b