Skip to content
Accnomos
Sign in

Meta models in Accnomos

Which Meta model does which job, where you see it in the app, how it is called, what code checks after it, and how it scored on our test sets.

The rule everywhere: AI proposes, you approve, code calculates. A model reads or words things; code works out every amount in pence and checks what the model said; nothing is posted or sent until you approve it. The test sets are made-up documents, so real phone photos will score lower. Checked against the code on 29 Oct 2026.

On our live server the Meta Model API (Muse Spark 1.3) answers first and a local fallback model answers last. Our live server uses Muse Spark 1.3 in Contributor mode for the demo and testing. In Contributor mode Meta may use the content sent to it to improve its models. The demo uses only fictional sample data; anyone trying it with their own books is served the same way, so please test with documents you are happy to share. Before real customers use it, the server switches to standard mode. With your own Meta key (key mode) the app asks for standard Muse Spark 1.3.

  • Muse Spark 1.3

    muse-spark-1.3-contributor (live, demo and testing) · muse-spark-1.3 (your own key)

    Reads documents and CIS and property paperwork; answers Ask; words the summaries; bank, inbox, reminder and duplicate suggestions

    Meta Model API: 1st on our live server (Contributor mode, demo and testing), or standard with your own key

    Measured: document type 96.5 %; clean documents 95 %; messy documents 96 %

  • Muse Glimmer 30B

    muse-glimmer:30b

    The same jobs as Muse Spark, as the local fallback, or for everything in local mode; most of our measured results

    A local fallback model, last on our live server; or your own machine

    Measured: clean documents 85 %; messy documents 80 %; CIS 92.5 %; property 99.1 %; Ask figures 32/32; bank coding 91.6 %

  • Muse Spark Open

    meta-models/muse-spark-open

    The Verifier: an independent second reading of each document

    Hosted open weights; switched off on the live server for now

    Measured: clean documents 91 %; messy documents 93 %

  • Muse Voice Transcribe

    muse-voice-transcribe-1.0

    Speech to text for spoken questions and instructions on Ask

    Meta Model API first, then a local fallback model

    Measured: word error rate 12.7 %

  • Segment Anything 3.1

    sam-3.1

    Finds and crops the receipt in a phone photo before it is read

    Meta Model API (/responses), when the server owner switches it on

    Measured: crop overlap 0.901

  • Muse Image

    muse-image-1.0

    Made the brand artwork and the demo video title art; not in the running app

    Meta Model API, from a build script

    Measured: Not measured yet

Muse Spark 1.3

muse-spark-1.3-contributor (live, demo and testing) · muse-spark-1.3 (your own key)

Meta's hosted Muse model on the Meta Model API: it reads images and text and replies in strict JSON. The first choice on our live server, in Contributor mode for the demo and testing.

Jobs, and where you see them

  • Says what kind of document each upload is (receipt, bill, sales invoice, credit note, statement) before it is read; you confirm or change it — Upload
  • Reads receipts, invoices, credit notes and emailed bills into a strict list (supplier, VAT number, date, lines, net, VAT, total) and suggests the category — Review
  • The Verifier's second reading in key mode and on the server chain — Review card: “Verifier agrees”
  • Reads CIS payment and deduction statements and reverse-charge invoices — CIS
  • Reads letting agents' statements, property bills, insurance schedules and mortgage statements — Property
  • Answers questions about your books, and turns typed or spoken instructions into an action card — Ask
  • Words the summaries: Home, month-end close, spending, cash flow and the day plan — Home, Month end, Spending, Reports › Cash flow, Jobs
  • Suggests an account and VAT code for a bank line no rule matched — Bank › Reconcile
  • Sorts email threads and drafts a reply and to-dos; drafts payment reminders — Home inbox, Invoices
  • Gives a view on possible duplicate contacts and on differently named items that may be the same product — Contacts, Prices
  • Answers the website's “Ask about Accnomos” chat from our public notes only — Public website

How it is called

  • On our live server, with no AI chosen, calls go in this order: 1st Muse Spark 1.3 on the Meta Model API (Contributor mode for the demo and testing); last, a local fallback model (Muse Glimmer 30B, open weights on our own server). A timeout, outage or rate limit moves the call to the next; a provider that keeps failing is skipped for a while.
  • Key mode: the Meta Model API (api.meta.ai/v1, /chat/completions) with your own Meta key, which stays on your device and is never logged.
  • Our live server uses Muse Spark 1.3 in Contributor mode for the demo and testing. In Contributor mode Meta may use the content sent to it to improve its models. The demo uses only fictional sample data; anyone trying it with their own books is served the same way, so please test with documents you are happy to share. Before real customers use it, the server switches to standard mode. With your own Meta key (key mode) the app asks for standard Muse Spark 1.3.

What code checks after it

  • AI proposes, you approve, code calculates: the model returns strict JSON or words only; every amount is worked out in pence by code, and nothing is posted, sent or kept until you approve it.
  • Readings are checked by code: lines and totals add up, VAT maths and rates, VAT number shape, duplicates, dates. Anything that fails waits in Needs a look.
  • Summaries and answers are checked figure by figure against facts code built from your books (the KnownNumbers check). An invented figure, or one from the wrong period, means code's own wording is shown instead, with the reason.

Measured

  • What kind of document is this (the upload flow's first step: receipt, bill, sales invoice, statement…): right type 96.5 %; 98.6 % answered by the model, the rest by code (bank CSVs) (1 Nov 2026 · 434 documents) Report: evals/runs/classify/report.md
  • Receipts and invoices, clean (reader and Verifier): every field right 95 %; second reading caught 0 wrong reading(s), missed 3, raised 2 false alarm(s) (4 Nov 2026 · 200 documents) Report: evals/runs/spark13_200/report.md
  • Receipts and invoices, messy photos (reader and Verifier): every field right 96 %; second reading caught 0 wrong reading(s), missed 3, raised 0 false alarm(s) (4 Nov 2026 · 200 documents) Report: evals/runs/spark13_messy_200/report.md
  • Every other job: Run live on our server, not scored against an answer key: Ask (typed and spoken), the Home summary, the goals note and the upload's document type and reading were each answered by Muse Spark 1.3 in a demo session. CIS, property, bank, inbox and reminder jobs are measured on Muse Glimmer 30B below. (4 Nov 2026 · live checks on the live server) Report: evals/runs/live_checks_2026-11-04/report.md

Muse Glimmer 30B

muse-glimmer:30b

Meta's open-weights Muse model: it reads images and text and replies in strict JSON. We run the weights ourselves, as the local fallback.

Jobs, and where you see them

  • Every Muse Spark job above, when the Meta Model API cannot answer: reading documents, CIS and property paperwork, Ask and action cards, the summaries, bank suggestions, inbox triage, reminders, duplicates and item matches — Review, CIS, Property, Ask, Home, Month end, Spending, Bank, Invoices, Contacts, Prices
  • All of them, when you choose local mode in Settings, including the Verifier's second reading — the same screens

How it is called

  • On our live server, with no AI chosen, calls go in this order: 1st Muse Spark 1.3 on the Meta Model API (Contributor mode for the demo and testing); last, a local fallback model (Muse Glimmer 30B, open weights on our own server). A timeout, outage or rate limit moves the call to the next; a provider that keeps failing is skipped for a while.
  • The code also supports Muse Glimmer 30B on a hosted GPU as an optional extra provider in that order; it is not what runs live.
  • Local mode: your own machine running Muse Glimmer, chosen in Settings. Documents then go only there.

What code checks after it

  • AI proposes, you approve, code calculates: the model returns strict JSON or words only; every amount is worked out in pence by code, and nothing is posted, sent or kept until you approve it.
  • Readings are checked by code: lines and totals add up, VAT maths and rates, VAT number shape, duplicates, dates. Anything that fails waits in Needs a look.
  • Summaries and answers are checked figure by figure against facts code built from your books (the KnownNumbers check). An invented figure, or one from the wrong period, means code's own wording is shown instead, with the reason.

Measured

Muse Spark Open

meta-models/muse-spark-open

The open-weights Muse Spark, on a hosted endpoint the server owner can switch on. A different model from the first reader, so the second reading is independent. It is off on the live server for now: the Verifier runs on the same Meta chain as the reader (Muse Spark, then Muse Glimmer). The results below are from earlier runs when it was on.

Jobs, and where you see them

  • The Verifier: a second reading of each document with a different prompt and schema that never sees the first answer — Review card: “Verifier agrees” or “Needs a look”

How it is called

  • Key mode and the server chain: only when the server owner switches it on (off for now); otherwise the chain's providers in order.
  • Local mode never uses it: your documents stay on your machine.

What code checks after it

  • The model never merges or picks values: code compares six fields (supplier, VAT number, date, net, VAT, gross). A disagreement sends the card to Needs a look; it never changes a figure.

Measured

  • Receipts and invoices, clean (reader and Verifier; earlier run): every field right 91 %; second reading caught 1 wrong reading(s), missed 3, raised 8 false alarm(s) (27 Oct 2026 · 200 documents) Report: evals/runs/spark_open_200/report.md
  • Receipts and invoices, messy photos (reader and Verifier; earlier run): every field right 93 %; second reading caught 0 wrong reading(s), missed 3, raised 3 false alarm(s) (27 Oct 2026 · 200 documents) Report: evals/runs/spark_open_messy_200/report.md

Muse Voice Transcribe

muse-voice-transcribe-1.0

Meta's speech-to-text model on the Meta Model API.

Jobs, and where you see them

  • Turns a spoken question or instruction (overtime, a quote, a shopping list) into text for Ask and action cards — Ask: the microphone button, on the web and Android

How it is called

  • 1st: the Meta Model API (/asr/transcribe), with your key in key mode, else the server's.
  • 2nd: a local fallback model, if the first fails. If both fail you see why, and can type the question instead.
  • The audio is converted in a temporary folder that is deleted at once; it is not kept.

What code checks after it

  • The text goes through the same path as a typed question: figures checked against your books, action cards priced in pence by code and waiting for Approve.
  • The answer, the action card and the audit log say which service heard you (“Transcribed by Muse Voice Transcribe”).

Measured

  • Spoken questions to text: word error rate 12.7 %; spoken numbers heard right 50 %, so a spoken card reads its figures back in words before you approve; 0 failures (29 Oct 2026 · 14 recordings) Report: evals/runs/stt/report.md
  • Used live: Used from the Android app on a real phone by the owner, and in our live checks: two spoken questions (synthetic speech) transcribed by Muse Voice Transcribe on the live server, then answered by Muse Spark 1.3. (4 Nov 2026 · live server) Report: evals/runs/live_checks_2026-11-04/report.md

Segment Anything 3.1

sam-3.1

Meta's Segment Anything model: finds an object in a photo from a word.

Jobs, and where you see them

  • Finds and crops the receipt or invoice in a messy phone photo before it is read — Review (photo uploads)

How it is called

  • The Meta Model API (/responses), only when the server owner switches it on (SAM_CROP=1): with your own Meta key in key mode, else the server's. Not when you choose a local model first or local only in Settings › AI.
  • Otherwise, and whenever it is unsure, errors or times out, a plain-code crop runs instead (no AI).

What code checks after it

  • Code accepts the crop only with a confident box (score at least 0.5) covering enough of the photo; any doubt or error keeps the original photo.

Measured

  • Crop, live on the Meta Model API: a box for 20 of 20 photos; mean overlap 0.901 against 0.88 for the plain-code crop; every field read right 19/20 after the SAM crop and 19/20 after the code crop; median 1.5 s a call (4 Nov 2026 · 20 synthetic scenes) Report: evals/runs/sam_crop_live/report.md
  • Plain-code crop (the fallback): mean overlap 0.88 against 0.48 for the uncropped photo (27 Oct 2026 · 20 synthetic scenes) Report: evals/REPORT_crop_code_20_2026-10-27.md

Muse Image

muse-image-1.0

Meta's image model. Used once, at build time, not by the running app.

Jobs, and where you see them

  • Made the brand artwork: the hero illustration, the app icon and the store graphic, also used for the demo video's title art and thumbnail — Website, Android icon, demo video

How it is called

  • The Meta Model API, from a build script (tools/brand_muse.py). The app never calls it.

What code checks after it

  • No figures involved; the images were chosen by a person.

Measured

  • Images made: 12 images, USD 0.12 in all (USD 0.01 each) (build time · 12 images)

No other Meta model is called. Every figure above is read from the committed eval runs when this site is built. See also how each feature works.