Sign in
Pending benchmarksTracking since Nov 3, 2026

Is Smaug-Audio-Decisions-0.7B state of the art?Pending benchmarks.Announced Nov 3, 2026; no independent results yet.

Abacus.AI's 744M audio decision model: it answers a typed question about a clip of up to 30 seconds with a calibrated probability per option in one forward pass, on a GPU, a Mac or offline on an iPhone. Fine-tuned from Google's EmbeddingGemma 2; open weights under Apache 2.0, with int8 Core ML builds.

Announced facts

The details so far

Lab
Abacus.AI
Announced
Nov 3, 2026
API
Not yet · No hosted API stated; runs locally, including an iOS app in abacusai/Smaug-Decisions-Mobile
Open weights
Published · Hugging Face abacusai/Smaug-Audio-Decisions-0.7B: float32 Transformers weights and int8 Core ML models; not gated
Parameters
744M (744,371,512 stored parameters, audio tower included)
Type
Audio decision model (audio classification): a probability per option, softmax for one answer or sigmoid for select-all; no text generation
Base model
google/embeddinggemma-2 (fine-tuned, with a linear head at a marker token per option)
Input
Audio up to 30 s (16 kHz mono) plus a question with 2 or more options; text and images are accepted, but the card says text decisions are close to chance
Cascade
Designed to sit in front of abacusai/Smaug-Decisions (4B) and hand low-confidence questions to it
License
Apache 2.0

Tags marked unsure are not stated on the lab's own pages.

Next steps

What would let us call it

Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.

  • Publicly announced with a source
  • Usable by the public (API or weights)
  • Listed on an independent board we track
  • Enough results for a rank in a category
  • Holds across two snapshots (crowns only)
Self-reported

What the lab claims

Self-reported. Shown for context; not used in any verdict.

Lab claim: on top of the leaderboard (Bindu Reddy's post; no leaderboard named).

Lab claim: 53.0 audio skill on Abacus's own 6,671-item Earshot benchmark, against 50.5 for Gemini 3.8 Flash and 43.6 for AudioJev, with 0.086 calibration error.

Lab claim: 9.7 ms per decision on an H100 and about 0.3 s offline on an iPhone 14 Pro Max; as a cascade in front of Smaug-Decisions it matches that model's 60.1 at 130 ms instead of 210 ms.

Receipts

Sources

LABBindu Reddy (Abacus.AI CEO) announcement on Xx.com/bindureddy/status/2117659234448154646 · Nov 3, 2026REPOabacusai/Smaug-Audio-Decisions-0.7B on Hugging Facehuggingface.co/abacusai/Smaug-Audio-Decisions-0.7B

Tracking continues until an independent board lists it.