PurpleMIST-Mini-1.0

PurpleMIST-Mini-1.0 Highlights

PurpleMIST-Mini-1.0 is the smallest System One decision model in the PurpleMIST family. Give it a piece of unstructured state (a message, ticket, record or transcript) and any number of typed questions about it. It returns a calibrated probability distribution for every answer, in one forward pass. It uses the same interface as PurpleMIST-Flash-1.0, at under a quarter of the size.

• 1.9B parameters. About 7.5 GB in memory, in fp32.
• 0.601 zero-shot accuracy in English on LocalLLaMA/typed-decisions (KL 0.280, Brier 0.156). That is above models up to four times its size: Bongard-mini (7.5B, 0.594) and Jeff-Gemma4-E2B (4.6B, 0.561). It is also well above Jeff-Qwen3.5-2B (2.2B, 0.511) and the input-blind Prior (0.470). It is below OpenDecider-small, a model twice its size (4B, 0.648 measured by us).
• Calibrated. ECE is 0.058 in English and 0.050 across eight African languages.
• One pass per state. All questions about a state share one sequence, and any number of options works.
• For African languages, use Flash. Mini reaches 0.416 on the translated African Typed Decisions cases, against 0.668 for Flash. A 2B backbone carries much less of these languages.

Model Overview

• Type: decision model. It scores options and does not generate text.
• Backbone: the text model of Qwen/Qwen3.5-2B-Base, with the vision tower and language-model head removed. It has 24 layers (18 Gated DeltaNet linear-attention and 6 full attention), hidden size 2048 and about 1.9B parameters.
• Head: a pointer head. Each option's logit is a scaled dot product between projections (2048 → 1024) of the answer position and the option's own line. The head has 4.2M parameters.
• Fine-tuning: full fine-tune of every weight. Flash uses LoRA instead.
• Precision: run it in fp32. Weights are stored in bf16, but the model trained with fp32 weights and bf16 matrix multiplies. Running it entirely in bf16 costs about 3 points (0.568 in English, against 0.601). Decider reads inference_dtype: float32 from purplemist_config.json and loads fp32 automatically.
• Question types:

Type	You give	You get back
choice	instructions and named options (criteria: key → description)	{"choice": key, "probabilities": {key: p}}
noul	a yes/no statement	{"noul": p_yes}
score	instructions and ordered level descriptions (lowest first)	{"score": expected level, "probabilities": {"0": p, ...}}

• Context: 2,048 tokens per pass. If all the questions do not fit alongside the state, they are split across several passes automatically.

Quickstart

pip install "transformers>=5.19" torch huggingface_hub
pip install flash-linear-attention causal-conv1d     # optional: fast kernels for the linear-attention layers

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's 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"]},
}
print(d.decide(state, questions))
# {"team": {"choice": ..., "probabilities": {...}}, "refund": {"noul": ...}, "urgency": {"score": ..., "probabilities": {...}}}

state can be a string or any JSON-serialisable object.

Serving over HTTP

serve.py runs the model as a small local server with the System One request shape (POST /v1/systemone), so clients built for that API can call it. It needs only the packages from the quickstart.

hf download olaverse/PurpleMIST-Mini-1.0 serve.py purplemist.py --local-dir purplemist
python purplemist/serve.py --model olaverse/PurpleMIST-Mini-1.0 --port 8000

curl -s localhost:8000/v1/systemone -H 'content-type: application/json' -d '{
  "state": {"message": "I was charged twice, please refund me"},
  "questions": {
    "team":   {"type": "choice", "instructions": "Which team?", "criteria": {"billing": "payments", "tech": "app problems"}},
    "refund": {"type": "noul", "instructions": "The customer asks for money back."}}}'
# {"model": "...", "answers": {"team": {"choice": ..., "probabilities": {...}}, "refund": {"noul": ...}}, "latency_ms": ...}

The server also offers GET /health and GET /v1/models. It handles one request at a time. --max-len raises the 2,048-token limit per pass, for very long states or questions with hundreds of options. The model was trained on 2,048 tokens, so answers on much longer inputs may be weaker. --no-truncate never shortens a request: anything that does not fit is declined with HTTP 422, for benchmarks whose rules forbid truncation. Ollama and llama.cpp cannot run this model: they serve text generators, and PurpleMIST scores options with its own head instead of generating text.

Evaluation

All models were run through one harness. Each request contains a whole case: the state and all of its questions. Every answer is turned into a probability vector over the same options, and every model is scored by the same code. LLMs were prompted for JSON probabilities at temperature 0 with reasoning off.

LocalLLaMA/typed-decisions (English, zero-shot)

These are the zero-shot rows from the benchmark card, with our measured rows added. The test split has 400 cases and 2,000 decisions.

Model	Size	Accuracy ↑	KL ↓	Brier ↓	ECE ↓	Source
meraGPT Decider 1	undisclosed	0.768	0.096	0.052	0.180	benchmark card
Liquid AI d1	undisclosed	0.742	0.475	0.155	0.124	benchmark card
TypeSafe Jev 1.13.0	undisclosed	0.727	1.442	0.148	0.144	benchmark card
PurpleMIST-Flash-1.0	7.9B	0.720	0.189	0.098	0.056	measured by us
DeepSeek-V4-Pro (prompted)	1.6T	0.720	1.359	0.185	0.108	measured by us
Featherless Simple Jev	35B (3B active)	0.716	0.488	0.176	–	benchmark card
Qwen3.8-Flash (prompted)	undisclosed	0.712	0.781	0.172	0.078	measured by us
prima-ratio + 12B	12B	0.702	0.564	0.234	0.146	benchmark card (self-reported)
OpenDecider-small	4B	0.671	0.211	0.117	–	benchmark card (self-reported)
OpenDecider-small	4B	0.648	0.224	0.126	0.059	measured by us
PurpleMIST-Mini-1.0	1.9B	0.601	0.280	0.156	0.058	measured by us
Bongard-mini	7.5B	0.594	0.256	0.132	0.067	benchmark card (self-reported)
Jeff-Gemma4-E2B	4.6B	0.561	0.403	0.219	0.188	benchmark card
Jeff-Qwen3.5-2B	2.2B	0.511	0.460	0.237	0.203	benchmark card
Jeff-Qwen3.5-0.8B	0.85B	0.483	0.679	0.313	0.251	benchmark card
Prior (ignores the input)	–	0.470	0.347	0.189	0.088	benchmark card

Sizes are total parameters; "undisclosed" means the provider does not publish one. Submitters compute ECE differently, so compare KL and Brier rather than ECE.

African Typed Decisions

Model	Size	English original	African, translated	English → African drop	African, native	KL ↓	Brier ↓	ECE ↓
DeepSeek-V4-Pro (prompted)	1.6T	0.720	0.669	0.051	0.955 †	1.125	0.199	0.121
PurpleMIST-Flash-1.0	7.9B	0.720	0.668	0.053	0.932 ‡	0.214	0.115	0.049
Qwen3.8-Flash (prompted)	undisclosed	0.712	0.643	0.070	0.916 †	0.862	0.216	0.132
OpenDecider-small	4B	0.648	0.477	0.171	0.635	0.365	0.199	0.027
PurpleMIST-Mini-1.0	1.9B	0.601	0.416	0.185	0.769 ‡	0.399	0.221	0.050
laya-multilingual	0.32B	0.348	0.322	0.026	0.280	5.053	0.545	0.360

KL, Brier and ECE are on the translated set: 3,010 cases, 15,050 decisions.

† DeepSeek-V4-Pro wrote the Nigerian-language native cases, and Qwen3.8-Flash wrote or checked all of them, so both are scored partly against their own answers on that column. ‡ PurpleMIST was trained on native cases generated by the same pipeline. They were separate cases from those in the benchmark.

By language (accuracy, translated cases):

Model	Amharic	Hausa	Igbo	Pidgin	Somali	Swahili	Yorùbá	isiZulu
PurpleMIST-Flash-1.0	0.649	0.649	0.654	0.710	0.683	0.690	0.644	0.671
OpenDecider-small	0.491	0.421	0.473	0.619	0.443	0.510	0.439	0.456
PurpleMIST-Mini-1.0	0.330	0.414	0.422	0.589	0.371	0.418	0.420	0.406
laya-multilingual	0.334	0.326	0.317	0.351	0.299	0.311	0.352	0.299

Mini does best on Nigerian Pidgin, which is closest to English, and worst on Amharic, the one language here not written in Latin script.

Training

• Data: the same training data as Flash. That is 300,000 states sampled from a 1.71M-decision set, with every African state included. The sources are:

  • public System One and preference data: tasksource, Open-Jev train, HelpSteer2, jev-decisions-clean50k, two synthetic sets and Najd development;
  • 86,134 African decisions, written natively in or translated into eight languages.

  Every source allows commercial use. LocalLLaMA/typed-decisions train was not used, which keeps the English result zero-shot.

• Objective: soft-target cross-entropy against each question's reference distribution. Options are shuffled during training.

• Run: full fine-tune for 6,007 steps, with fp32 master weights and bf16 autocast. The learning rate was 2e-5 and weight decay 0.01. Sequences were up to 2,048 tokens.

• Calibration: one temperature per question type, fitted in fp32 on 2,751 states (5,345 questions) that training never used. The states are spread across eight sources, each weighted equally. The temperatures are large: choice 24.5, noul 33.5, score 23.5. Mini's raw scores are very sharp, and the temperatures bring them back to calibrated probabilities. Decider applies them for you. If you read raw logits yourself, divide by these.

Limitations

• African languages: 0.33–0.59 on translated cases. Use PurpleMIST-Flash-1.0 for these languages.
• Gold comes from LLM teachers. Reference answers were generated by language models, so a score measures agreement with those teachers, not ground truth.
• Text only. Images in the state are not read.
• bf16: running in bf16 lowers accuracy. Keep the default fp32.
• eval/ files: these come from the training run's evaluation, in bf16 and before recalibration. The numbers on this card replace them.
• Score questions: score returns the expected level. Use probabilities if you need the most likely level.

Citation

@misc{purplemist-mini-1.0,
  title  = {PurpleMIST-Mini-1.0: a small calibrated System One decision model},
  author = {Olaverse},
  year   = {2026},
  url    = {https://huggingface.co/olaverse/PurpleMIST-Mini-1.0}
}