Aplomb 1: A 5.3B-Parameter Decision Model with 1M Context and Four-Modality Input

EmpirioLabs has released an open-weights decision model called Aplomb 1. With only 5.3 billion parameters, it stands out for several distinctive capabilities: a 1 million token context window, simultaneous processing of text, images, video, and audio in a single request, and probability scores returned for tool-call parameters. Built on Qwen3.5-4B Base, it tops the official Decision Index leaderboard among 4B-class models.

A Model Built Specifically for "Decision-Making"

Unlike most general-purpose chat models, Aplomb 1 positions itself as a "decision model." Its core value lies not in generating fluent long-form text, but in quickly delivering probabilistic judgments — determining which tool to invoke, what parameters to fill in, and whether an input actually contains an answer.

The official example illustrates this well: given the sentence "Order B-44120 arrived with a cracked screen, I want a refund" and three available tools, the model assigns a probability of 0.969 to selecting issue_refund, 0.993 to the reason "damaged", and 0.761 to full_refund being true.

This design — returning probabilities for every tool, every enum value, and every boolean parameter — lets agent systems route accordingly: high-confidence calls execute immediately, while uncertain ones get escalated to a larger model. It's a pragmatic layered architecture: use a small model for upfront judgment, and reserve the expensive large model for genuinely hard problems.

This "decision model" positioning addresses a long-standing efficiency bottleneck in AI agent systems. In a complete agent workflow, many steps don't require complex reasoning — they're just classification, routing, or parameter extraction: determining which category a user's intent falls into, extracting structured fields from text, or confirming whether a condition holds. Routing all of these through a massive flagship model is both slow and expensive. Probabilistic output further amplifies this advantage: traditional classifiers return a single discrete result, while Aplomb 1's confidence distributions enable "soft routing" downstream — requests are only escalated to a more expensive model when confidence falls below a threshold, dynamically balancing cost against accuracy.

1M Context and Inference Speed

Aplomb 1's context window has been extended from Qwen3.5-4B's native 262K to 1 million tokens, supporting mixed inputs of text, JSON, images, video, and audio within a single request. According to the official description, this is currently the only known decision model that can simultaneously handle 1M tokens, accept four modalities, and return probabilities for tool parameters.

Speed is another key highlight. The company's proprietary inference runtime features a long-context mode that delivers decisions in approximately 3 seconds when processing a 1M token document — while a full sequential read takes around 111 seconds. In long-context tests, the model answered all 525 decision questions correctly.

For short queries, model-side computation takes roughly 15 milliseconds, with end-to-end latency around 200 milliseconds. API pricing is $0.02 per million input tokens, with output tokens free, and zero data retention (ZDR) enabled by default. It's worth noting that the 3-second long-context speed applies to the official API's inference runtime; the open-weights version reads every token.

Benchmark Results and Data Disclosure

On Decision Index 0.2.1, Aplomb 1 scored 44.86 points using official tooling, ranking first among published 4B-class models. Across 38 benchmarks, it achieved the highest scores among sub-5.3B models on 8 tasks, including MMLU-Pro, GPQA Diamond, and BBH — three of the more credible tests in the field.

Other results include 77.5% on JevBench Hard and an average zero-shot intent recognition accuracy of 75% across 51 languages on the MASSIVE dataset. Multilingual capability is a meaningful plus for agent systems intended for global deployment.

Noteworthy is the team's proactive data disclosure: training data includes the public training splits of WinoGrande and ContractNLI — both of which appear in the 38-benchmark index. This means scores on those two should be taken with a grain of salt. This transparency is refreshing in an environment where benchmark gaming is a widespread concern, and it actually increases overall credibility.

MMLU-Pro, GPQA Diamond, and BBH are widely regarded as among the most demanding benchmarks in current language model evaluation. MMLU-Pro upgrades the original MMLU by expanding answer choices from 4 to 10 and introducing more reasoning-heavy professional questions, making it harder to game through memorization. GPQA Diamond targets PhD-level science questions in physics, chemistry, and biology — problems that even domain experts struggle to answer reliably — making it a credible measure of deep reasoning. BBH (BIG-Bench Hard) selects 23 tasks from BIG-Bench on which leading models at the time scored significantly below random baselines, focusing on logical reasoning and symbolic manipulation that can't easily be solved through data memorization. Aplomb 1 leading its peer class on all three is the primary basis for treating its results as relatively meaningful.

Deployment Requirements and Licensing

Model weights run in bf16 precision and require approximately 12 GB of VRAM with the reference script — meaning a single consumer-grade GPU is sufficient. This sets a notably low deployment barrier. The model is built on Qwen3.5-4B and integrates the audio encoder from Qwen3-Omni-30B-A3B-Instruct, both of which are licensed under Apache-2.0.

However, licensing requires attention: Aplomb 1 uses EmpirioLabs' proprietary Model License, which is free for research, evaluation, personal use, and internal use by companies with annual revenue under $1 million. This is a common "source-available" rather than fully open-source licensing model — large commercial deployments will require separate negotiation.

"Open-weights" and "open-source" are two terms that are often conflated but mean different things. Open-source typically requires releasing model weights, training code, and datasets, with permission for arbitrary use and modification. Open-weights simply means the model parameter files are publicly available — training details and data sources may remain undisclosed, and licensing can include commercial restrictions. Aplomb 1's EmpirioLabs Model License is a textbook example of this "source-available" model: weights can be downloaded and run locally, but commercial use is gated by a revenue threshold. This licensing structure has become increasingly common in recent model releases — Llama, Mistral, and others follow similar approaches — and developers should carefully review the terms before integrating any such model into a commercial product.

Takeaway: A Pragmatic Small-Model Strategy

Aplomb 1 represents a product philosophy worth watching — rather than competing with giant models on parameter count, it sharpens a small model into a specialized tool for specific tasks: decision-making, tool invocation, and long-document judgment. The combination of probabilistic output, ultra-long context, four-modality input, and low VRAM requirements makes it genuinely attractive to developers building layered agent systems.

Of course, these benchmark scores and speed figures come from the vendor itself, and independent third-party verification is still needed. Interested developers can download the weights on Hugging Face or try the model directly in the official Playground.