Aplomb 1: Decision Model with Audio and 1M Context
EmpirioLabs released Aplomb 1, a 5.3B decision model that reads text, JSON, images, video and audio with 1M token context at $0.02 per million tokens.
👁 116 reads
EmpirioLabs released Aplomb 1, an open decision model with 5.3 billion parameters that reads text, JSON, images, video and audio in a single request, accepts up to 1 million tokens of context, and returns calibrated probability distributions instead of generating text. It’s the first model in this category to support audio, and has the longest context among open decision models. We covered the entire category and tested other models in our guide AI Decision Models.
As of October 7, 2026
- Release: September 29, 2026 (date in EmpirioLabs documentation), model card updated October 5 and 6.
- Model: 5.3 billion parameters, based on Qwen3.5-4B, text + JSON + image + video + audio, 1 million token context.
- Weights: on Hugging Face behind a gate requiring acceptance of EmpirioLabs Model License 1.0 (not Apache-2.0).
- API: $0.02 per million input tokens, output free, requires key.
- Results: all provided by the producer, no independent verification yet.
What Aplomb 1 Is
A decision model doesn’t generate sentences. It receives a state (e.g. customer inquiry, document, photo, recording) and a set of questions with closed answer lists, and returns probability distributions for each option. Aplomb 1 supports four question types:
- noul - yes/no with probability,
- choice - selection from a list of 2 to 255 options,
- score - rating on an ordered scale from 2 to 10 levels,
- tool - function selection from a list along with probability distributions for arguments.
The first three types are the standard popularized by Jev from TypeSafe and repeated by Cloudflare Clef and Strands Decider from AWS. The fourth, tool selection with arguments, is new in this category.
Key Numbers
| Parameter | Aplomb 1 |
|---|---|
| Size | 5.3 billion parameters (based on Qwen3.5-4B) |
| Input | text, JSON, image, video, audio |
| Context | up to 1 million tokens |
| Single request limits | up to 16 images, 8 audio recordings, 4 videos |
| Questions per request | 128 according to model card, 512 according to API documentation |
| Memory for local deployment | approx. 12 GB (bf16, according to authors) |
| API price | $0.02 per million input tokens, output $0 |
| Decision time in API (according to producer) | approx. 15 ms short question, 35 ms with image, approx. 3 s at 1 million tokens |
The discrepancy in questions per request (128 or 512) comes from two different producer documents. We haven’t verified which limit applies in the API.
Results According to EmpirioLabs
On the Decision Index 0.2.1 test, Aplomb 1 scores 44.86 points. The producer notes they ran the official set themselves on October 5, and this is not an entry on the published leaderboard.
| Decision Index Category | Aplomb 1 | JPT-4B |
|---|---|---|
| Overall score | 44.86 | 43.04 |
| Knowledge and reasoning | 29.7 | 28.7 |
| Language understanding | 48.3 | 52.5 |
| Search and classification | 49.5 | 45.0 |
| Tools and automation | 62.4 | 57.2 |
| Art and taste | 33.9 | 25.8 |
On images (500 examples each), Aplomb 1 is according to the model card on par with Intern-Decision-4B and OmniJev-4B: POPE 89.0, MMStar 68.6, AI2D 87.0, MMMU 57.2, HallusionBench 79.8. For video it reports 80.7 on Video-MME (short), and for audio 91.9% on vocal sound recognition and 78.6% on emotions in speech. Audio results were updated October 6 along with new audio weights.
In the feature table, EmpirioLabs compares Aplomb 1 with Jev 1.13 (64k token context, $0.042 per million tokens), Intern-Decision-4B (8192 tokens), and Decisions API from OpenAI (no published pricing), among others.
One thing worth noting: on the JevBench project GitHub there’s an issue requesting independent evaluation of Aplomb 1, where the author notes that public JevBench tasks were used in checkpoint selection. Results on public JevBench may therefore be inflated - this is information from the issue itself, not our measurement.
License: What to Watch For
This is the most important difference from Clef and Strands Decider, which have Apache-2.0. EmpirioLabs Model License 1.0 allows free use for:
- research, education, evaluation and personal use,
- commercial use within organizations with annual revenue below $1 million USD.
Hosted services, larger companies and distribution in a product require a commercial license. If you’re building SaaS that will run on Aplomb, expect negotiations or stick with the producer’s API.
Have We Tested It
Not yet. The weights are on Hugging Face behind a gate requiring license acceptance, and the API requires a key and account. The model would fit on our RTX 4090 (12 GB in bf16), so when we return to testing, we’ll run it through the same 30 inquiries in Polish on which Clef-flash and Strands Decider scored 30 out of 30. We’re particularly curious whether its declared calibration (“80% confidence means 80% accuracy”) will hold in Polish - with Strands Decider, probabilities in Polish were noticeably underestimated.
Who It’s For
- Yes: automations where decisions depend on recordings (e.g. evaluating helpline calls), long documents, or multiple images at once; agents that need to select tools along with arguments.
- Probably not: larger companies wanting to self-host the model without license discussions - Clef and Strands Decider on Apache-2.0 are simpler here.
Sources
- empiriolabsai/aplomb-1 on Hugging Face - model card, results, license, requirements
- Aplomb 1 Documentation - release date, pricing, endpoints, limits
- empiriolabsai/aplomb-1-omni - variant on Qwen3-Omni-30B-A3B
- JevBench, issue #195
Verified: October 7, 2026.
More on this from QUANTSLANT
ai tools·Google Antigravity: Models, Pricing, Limits & Claude 5.5Google Antigravity is an AI coding platform with desktop app, VS Code-based IDE and CLI. Claude Opus 5.5 and Sonnet 5.5 require paid Google AI Pro.