MIST-Extract-Flash-1.0
MIST-Extract at 2B, for documents at volume.
MIST-Extract-Flash-1.0 is the fast member of the MIST-Extract line. It takes the same JSON templates and the same text, photo or scan inputs as MIST-Extract-1.0, returns a confidence for every field, and processes 9.7 documents per second on a single GPU.
The problem it solves
Extraction workloads are often about volume: every receipt a field team photographs, every payment SMS a lender ingests, every form a back office receives. At that scale the cost and speed of each document matters as much as the last point of accuracy.
Flash keeps the MIST-Extract interface, its African-language strength and its per-field confidence at half the size, so you can run more documents on the same hardware and move up to the 4B model without changing your templates.
Highlights
Ahead of extractors twice its size
On 486 held-out English documents it gets 90.5% completely right, against 67.9% for NuExtract3 (4B) and 34.0% for Qwen3.5-4B. On English receipt photos it gets 99.9% of fields right.
Far ahead on African languages
0.701 field accuracy on MasakhaNER 2.0 named entities in 21 African languages, against 0.393 for NuExtract3. In Shona, Luo, isiXhosa and Fon it leads by 30 to 50 points.
Fast
9.7 documents per second on one RTX A6000 under vLLM, 1.7× NuExtract3 and 1.6× MIST-Extract-1.0.
Best on CORD receipts
On the CORD-v2 receipt test set it gets 71.0% of receipts fully right, the highest of every model we compared, the 4B MIST-Extract included.
Run it in a few lines
import sys
from huggingface_hub import snapshot_download
path = snapshot_download("olaverse/MIST-Extract-Flash-1.0")
sys.path.insert(0, path) # the repo ships mist_extract.py
from mist_extract import Extractor
x = Extractor.from_pretrained(path, device="cuda")
template = {"merchant": "verbatim-string", "date": "date", "total": "number"}
# Many documents in one batch
x.extract_batch([
{"template": template, "text": "Paid 2,500 naira at Mama Put on 3 May 2024."},
{"template": template, "image": "receipt.jpg"},
])Which MIST-Extract model?
Need the highest accuracy, especially on general documents and African languages? Use MIST-Extract-1.0.
| MIST-Extract-1.0 → | MIST-Extract-Flash-1.0THIS PAGE | |
|---|---|---|
| Parameters | 4B | 2B |
| English documents fully right (held-out) | 91.8% | 90.5% |
| Named entities, 21 African languages | 0.728 | 0.701 |
| Field accuracy, all held-out documents | 0.898 | 0.879 |
| Speed, one RTX A6000 | 6.0 docs/s | 9.7 docs/s |
| Confidence calibration error (ECE) | 0.008 | 0.022 |
| Best for | Top accuracy, general and African documents | Volume and receipts |
Best for
- High-volume receipt, bill and payment-message ingestion where cost per document matters.
- Receipt photos in African currencies and layouts.
- Prototyping an extraction pipeline before moving up to MIST-Extract-1.0.
- Smaller GPUs and shared inference servers.
More production models
All modelsPurpleMIST-Flash-1.0
typed-decisions 0.720 · African TD 0.668
PurpleMISTPurpleMIST-Mini-1.0
typed-decisions 0.601 · 1.9B
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