Smaug-Decisions

Abacus.AI's decision model for audio, images and text, a Gemma 4 E4B fine-tune with a 3.45 GB phone build. Apache 2.0.

In short. Smaug-Decisions answers typed questions about a sound clip, a picture or a piece of text, such as whether a speaker has finished talking or which genre is playing. It gives a probability for each possible answer instead of writing a reply. A tiny sibling answers audio questions in about 10 milliseconds, and a phone version runs fully offline.

Summary

Abacus.AI published Smaug-Decisions on Hugging Face on 26 October 2026. It is a decision model, so for each question it returns a probability per option in one forward pass instead of writing text. It is a full fine-tune of Google's Gemma 4 E4B-it, with 4B effective and 8.6B stored parameters, and it takes audio of up to 30 seconds, one image or text, plus a question with 2 to 20 options. A smaller sibling, Smaug-Audio-Decisions-0.7B, is a fine-tune of EmbeddingGemma 2 built mainly for audio questions, and Smaug-Decisions-LiteRT packs the main model into a 3.45 GB build for Google's LiteRT-LM on phones.

Abacus.AI reports 60.1 audio skill on Earshot, its own 6,671-item audio benchmark, against 59.4 for Gemini 4 Argon and 50.5 for Gemini 3.8 Flash through their APIs, with a calibration error of 0.069 against 0.123 and 0.204. On the public Jev Decision Index it reports 57.9 on images and 56.6 on text, the best Gemma 4 E4B fine-tune on both boards, though its own table shows Perplexity's 4B pplx-decider Flash ahead on text at 61.8. The 0.7B model scores 53.0 on Earshot at 9.7 ms per decision on an H100, and as a first tier that escalates unsure cases to the main model it matches 60.1 at 130 ms instead of 210 ms. The phone build keeps 59.4 on Earshot and takes about 0.6 s for a text decision on an iPhone 14 Pro Max, and its text index falls to 53.9.

The model ships with a small Python wrapper and needs Transformers 5.20 or newer. The main weights are bf16 and 17.2 GB, the 0.7B model's are about 3 GB, and the config sets a 131,072-token position limit that the card does not mention. Per request the card allows one audio clip of up to 30 seconds, where longer audio keeps its last 30 seconds, one image at 560 image tokens and up to 20 options. The license is Apache 2.0, the same as the Gemma 4 E4B-it base according to the card. In the 30 days to 11 November 2026, Hugging Face counted only 95 downloads for the main repository, 190 for the 0.7B model and 96 for the LiteRT build.

Good for

  • Audio decisions such as whether a speaker has finished their turn, who speaks next or which sound or genre is playing, with probabilities you can threshold
  • Phone apps and other offline uses, with the 3.45 GB LiteRT build or the 0.7B model
  • A two-tier setup where the 0.7B model answers confident audio questions in about 10 ms and the larger model handles the rest
  • One Apache 2.0 model for audio, image and text decisions

Watch out for

  • Earshot is Abacus.AI's own benchmark and the index rows are its own runs, and the API rivals received the audio and returned JSON probabilities, so treat the lead over Gemini as the developer's claim
  • On text the card's own table shows Perplexity's 4B decider ahead, 61.8 against 56.6, and index editions and settings differ between cards, so do not compare scores across cards
  • It answers typed questions with probabilities and cannot write text, and the card advises setting decision thresholds on your own data for high-stakes uses
  • Adoption is tiny so far, with 95 to 190 downloads per repository in 30 days, and the phone build gives up 2.1 to 3.0 points on the image index and 2.7 on text

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the Smaug-Audio-Decisions-0.7B variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
Smaug-Audio-Decisions-0.7B FP32 740M 1.7 GB 0.851 GB 0.511 GB Laptop or phone
Smaug-Decisions BF16 8.61B 20 GB 9.9 GB 5.94 GB Laptop or phone
Smaug-Decisions-LiteRT 8.61B 20 GB 9.9 GB 5.94 GB Laptop or phone

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is Smaug-Decisions?
Smaug-Decisions answers typed questions about a sound clip, a picture or a piece of text, such as whether a speaker has finished talking or which genre is playing. It gives a probability for each possible answer instead of writing a reply. A tiny sibling answers audio questions in about 10 milliseconds, and a phone version runs fully offline.
Can I use Smaug-Decisions commercially?
Yes. Smaug-Decisions is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Smaug-Decisions open source?
Smaug-Decisions is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run Smaug-Decisions on my own computer?
Yes. The smallest variant, Smaug-Audio-Decisions-0.7B, needs about 0.511 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was Smaug-Decisions released?
Smaug-Decisions was released by Abacus.AI on October 26, 2026.
How long a context window does Smaug-Decisions have?
Smaug-Decisions supports a context window of up to 128K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Liquid d1 Liquid AI 587M – 3.12B 128K LFM Open 1.0 (Conditional) 0.405 GB Laptop or phone
pplx-decider Perplexity 4.54B – 26.1B 256K Apache 2.0 (Permissive) 3.13 GB Laptop or phone
AutoTrust Decider AutoTrust AI Lab 8.95B – 36B 3B active 256K Apache 2.0 (Permissive) 6.18 GB Laptop or phone
Clef Cloudflare 9.41B – 27.4B 64K Apache 2.0 (Permissive) 6.49 GB Laptop or phone