PurpleMIST-Flash-1.0
Frontier-level decisions from a model 200× smaller.
PurpleMIST-Flash-1.0 is our frontier decision model. Give it a message, ticket, record or transcript and any number of typed questions about it, and it returns a calibrated probability for every possible answer in a single pass, in English and eight African languages.
The problem it solves
Most of what software decides about text is a set of small, typed questions: which team owns this ticket, is the customer asking for a refund, how urgent is it. Teams usually send each one to a huge general-purpose model, pay for every token, parse whatever comes back, and get a confident answer with no sense of how sure the model really is.
PurpleMIST answers all of those questions at once, scores every option directly instead of writing text, and tells you how likely each answer is. It does this as well as a 1.6-trillion-parameter model on African-language input, on a single 24 GB GPU.
See it decide
English: tickets, scams and invoices
Routing a damaged-order email, catching a phishing text and holding an invoice that doesn’t match its purchase order, with every question answered in one pass.
African languages: the same decisions, natively
A fake bank text in Yorùbá, a POS transfer complaint in Hausa and an M-Pesa scam in Kiswahili, with questions asked in the message’s own language.
The numbers
African languages: accuracy and probability quality
Across eight African languages it sits level with DeepSeek-V4-Pro on accuracy while landing far to the left on KL, meaning its probabilities are much closer to the reference.
English, zero-shot
On LocalLLaMA/typed-decisions, without training on its train split, it scores 0.720: 4th for accuracy and 2nd for KL and Brier among the zero-shot models listed.
Jev Decision Index, against models its size
Across 37 public benchmarks it scores 46.7 against a median of 41.9 for the 13 models of 7–10.5B parameters, and beats that median in four of the five areas, led by language understanding.
Highlights
Matches a model 200× its size
It ties DeepSeek-V4-Pro (1.6T) at 0.720 on LocalLLaMA/typed-decisions (English, zero-shot), scores 0.668 against its 0.669 on African Typed Decisions, and leads on Pidgin, Somali, Swahili and isiZulu.
Probabilities you can trust
On African Typed Decisions its answers sit about 5× closer to the reference distribution than DeepSeek-V4-Pro’s (KL 0.214 vs 1.125), with calibration error (ECE) of 0.049.
Built for African languages
Trained on 86,134 decisions in Amharic, Hausa, Igbo, Pidgin, Somali, Swahili, Yorùbá and isiZulu, alongside English.
Ahead of most models its size
On the Jev Decision Index public suite it scores 46.7, ahead of 8 of the 13 models of similar size and 4.8 points above their median, answering all 140,620 requests at a median 63 ms.
Run it in a few lines
import os, sys
from huggingface_hub import hf_hub_download
repo = "olaverse/PurpleMIST-Flash-1.0"
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)
# {"team": {"choice": ..., "probabilities": {...}},
# "refund": {"noul": ...}, "urgency": {"score": ..., "probabilities": {...}}}Which PurpleMIST model?
English-only workload on smaller hardware? PurpleMIST-Mini-1.0 has the same interface at 1.9B.
| PurpleMIST-Flash-1.0THIS PAGE | PurpleMIST-Mini-1.0 → | |
|---|---|---|
| Parameters | 7.9B | 1.9B |
| LocalLLaMA/typed-decisions (English, zero-shot) | 0.720 | 0.601 |
| African Typed Decisions (8 languages) | 0.668 | 0.416 |
| Hardware | One 24 GB GPU | About 7.5 GB in fp32 |
| Best for | African languages and top accuracy | English at lower cost |
Best for
- Routing and triaging support tickets, chats and emails in English and African languages.
- Compliance, moderation and KYC checks where you need a probability, not just a label.
- Replacing chains of prompted LLM calls with one fast pass that answers every question together.
- Self-hosted decision APIs on a single 24 GB GPU, served over the System One request shape with the same answer fields as Ollama’s decision endpoint.
- Long inputs such as full transcripts and case files, up to 65,536 tokens in one pass.
Use cases
Automated decisions
Route, flag and score tickets, chats and records with calibrated probabilities, in English and eight African languages, from one self-hosted model instead of a chain of LLM calls.
/ Extraction · DecisionsPayment messages
Pull the amount, sender, reference and date out of mobile-money and bank SMS, then decide in one pass whether it is genuine, which account it belongs to and whether anyone needs to look at it.
More production models
All modelsMIST-Extract-1.0
91.8% docs fully right · 4B
MIST-ExtractMIST-Extract-Flash-1.0
9.7 docs/s · 2B
LID-608lid-neural-608
0.867 acc · 140M
LID-608lid-lite-608
37 MB · 6,800 texts/s
DiacNetdiacnet-2.0
DER 0.0117 · 582M
DiacNetdiacnet-mini-2.0
DER 0.0141 · 300M
DiacNetdiactag-2.0
DER 0.0130 · 37.9M
MIST QGmist-qg-1.5b
0.955 keep-rate · 25 langs