PurpleMIST-Mini-1.0
PurpleMIST decisions in 1.9B parameters.
PurpleMIST-Mini-1.0 is the compact member of our frontier decision line. It takes the same state-and-questions request as PurpleMIST-Flash-1.0 and returns calibrated probabilities for every answer, at under a quarter of the size.
The problem it solves
Many decision workloads are English, high-volume and cost-sensitive: tagging every inbound email, scoring every lead, checking every form. They need calibrated answers, but not a 9B model on every request.
Mini keeps the PurpleMIST interface and calibration and fits on far smaller hardware, so you can start small and move up to Flash without changing a line of your integration.
The numbers
English, zero-shot
On LocalLLaMA/typed-decisions it scores 0.601 with a KL of 0.280, above Bongard-mini (7.5B), Jeff-Gemma4-E2B (4.6B), both Jeff-Qwen models and the input-blind prior.
Highlights
Punches above its size
At 0.601 zero-shot on LocalLLaMA/typed-decisions, it outscores Bongard-mini (7.5B) and Jeff-Gemma4-E2B (4.6B).
Calibrated out of the box
ECE 0.058 in English. Per-type temperatures are applied for you, so the probabilities are ready to threshold.
Same interface as Flash
Identical request and response shapes, Python class and HTTP server. Swap the model name to move up.
Small and fast
1.9B parameters and about 7.5 GB in memory in fp32, which it loads by default. Median 46 ms per request, with up to 32,768 tokens per pass.
Run it in a few lines
import os, sys
from huggingface_hub import hf_hub_download
repo = "olaverse/PurpleMIST-Mini-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") # loads in fp32 from the model config
state = {"channel": "email",
"message": "Hi, I was charged twice for the same order last night. Please refund the extra payment."}
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)Which PurpleMIST model?
Working with African languages, or need the highest accuracy? Use PurpleMIST-Flash-1.0.
| PurpleMIST-Flash-1.0 → | PurpleMIST-Mini-1.0THIS PAGE | |
|---|---|---|
| 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
- High-volume English classification and routing where cost per request matters.
- Prototyping a decision pipeline before scaling up to PurpleMIST-Flash-1.0.
- Smaller GPUs and shared inference servers.
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