# Mercury Decide: Inception's Fast AI Decision Model Explained > Source: > Published: 2026-10-21 00:00:00+00:00 # Mercury Decide: Inception's Fast AI Decision Model Explained Mercury Decide is a diffusion-based AI decision model that gives instant yes/no judgments with confidence scores, now with vision support. ## What is Mercury Decide? Mercury Decide is a diffusion-based decision model from Inception Labs, built to answer narrow, typed questions (yes/no, multiple choice, confidence scores) instead of generating free-form text. You feed it a state (text or, in its newest release, images) plus a set of structured questions, and it returns an answer with a probability score in a fraction of a second. It’s Inception’s answer to the kind of judge models that agent builders increasingly rely on to check whether a task succeeded, whether a ticket needs escalation, or whether a page actually loaded. ## TL;DR - **Mercury Decide** is a diffusion decision model that returns a structured answer and confidence score in one call, with no generated text involved. - It now supports **vision input** , letting it judge screenshots or images alongside text state, which matters for web agents and UI-checking tasks. - On third-party evaluations from Vals AI, it **matched or beat Gemini-style judge models** on claim verification and scored around 87% versus 74% on LegalBench. - It runs at roughly a **quarter of the cost** of comparable judge models and is available free on OpenRouter during its early release window. - Independent measurements put median judgment latency around **0.16 seconds for a single question** and about 0.4 seconds for 64 questions on the same context. - It pairs naturally with **Mercury 3** , Inception’s faster general-purpose diffusion LLM, as a cheap “system one” check wrapped around more expensive “system two” reasoning calls. - On the Artificial Analysis Decision Index, Mercury Decide ranked **third on hallucination resistance** at a 32% rate, a meaningful number for anyone who needs a model to say “not sure” instead of guessing. ## Other agents start typing. Remy starts asking. Scoping, trade-offs, edge cases — the real work. Before a line of code. ## What is a decision model, and why does it matter? A decision model is a narrower cousin of a general-purpose LLM. Instead of writing paragraphs, it takes in some context (a transcript, a screenshot, a log file) along with a specific, typed question and returns a compact judgment: true or false, a category, or a probability. No chain of reasoning gets printed out, no tokens get wasted on explanation. This matters because a huge share of what AI agents actually do isn’t creative generation, it’s constant small verification. Did the page load? Is this support ticket billing-related? Did the previous agent step actually complete? Running a full reasoning model for every one of these micro-checks is slow and expensive. A dedicated decision model is built to answer exactly these questions, fast and cheap, which is why they’re sometimes called “judge” models, following the pattern popularized by larger general models being used to grade or verify other models’ outputs. ## How does Mercury Decide actually work? Mercury Decide is built on diffusion, the same family of generative techniques behind most modern image and video models. Where a standard autoregressive LLM writes one token at a time, committing to each choice before moving to the next, a diffusion model starts with a masked, rough draft of the entire output and refines it across a handful of parallel passes. Because inference is bottlenecked by memory bandwidth rather than raw compute, doing more useful work per pass is where the speed gain comes from. For a decision model, that architecture is a natural fit. The output isn’t an essay, it’s a short structured answer, so there’s less to refine and less latency end to end. Mercury Decide takes in a “state” (which can now include images as well as text) along with a set of typed questions, and returns answers and confidence scores in one pass rather than generating and then parsing free text. Inception’s leadership has direct roots in diffusion research. CEO Stefano Ermon is credited as a co-inventor of some of the key diffusion techniques used in image and video generation, and the founding team has worked on related methods like DPO and flash attention. That lineage is part of why Inception pushed diffusion into language modeling rather than treating it as a side project. ## How does Mercury Decide perform against other judge models? Independent testing gives a clearer picture than vendor benchmarks alone. Vals AI ran Mercury Decide through the same evaluation suite used for Gemini-based judge models and found it already matching that performance on claim verification tasks. On LegalBench, a benchmark built around legal reasoning tasks, Mercury Decide scored around 87% compared to roughly 74% for the comparison judge, while reportedly costing about a quarter as much per call. On the Artificial Analysis Decision Index, Mercury Decide placed third for hallucination resistance, with a 32% hallucination rate. For a decision model, that number carries more weight than raw accuracy alone: a model that confidently guesses wrong is worse than one that flags uncertainty, especially in workflows where a low-confidence answer should trigger a human review instead of an automated action. On speed, Vals AI measured a median of about 0.16 seconds for a single judgment call and roughly 0.4 seconds when running 64 judgments against the same context in parallel. That parallel efficiency points to the model’s best use case: loading in one piece of context (a document, a conversation, a screenshot) and firing off many typed questions against it at once, rather than making repeated round trips. ## What does vision support add? The newest release of Mercury Decide adds image input to its state, meaning the model can now judge visual content alongside or instead of text. In practice that opens up checks like confirming a webpage rendered correctly, verifying a UI element appeared, or classifying the content of a screenshot, all as fast, typed yes/no or confidence-scored calls rather than routing images through a full vision-language model every time. This fits a broader pattern forming around agent design: cheap, fast “system one” checks wrapped around a single, more expensive “system two” call. An agent might use a full reasoning model like Mercury 3 to plan and execute a task, but lean on Mercury Decide at every step to confirm things are on track, escalating to a human or a bigger model only when confidence drops. ## How does Mercury Decide relate to Mercury 3? Mercury Decide doesn’t operate in isolation. It’s released alongside Mercury 3, Inception’s general-purpose diffusion LLM, which the company positions as the “thinking” counterpart. Mercury 3 reportedly generates over 1,000 tokens per second on Nvidia Blackwell hardware, which Inception describes as roughly eight times faster than GPT-6 Nano-class models and about 4.5 times faster than Claude Haiku 5.5, while scoring competitively with those same models on reasoning, instruction-following, and search-oriented benchmarks like GPQA Diamond. The pairing matters because it describes a full workflow: Mercury 3 handles the harder reasoning steps an agent needs, while Mercury Decide handles the constant stream of smaller verification questions around those steps. Both are built on the same diffusion foundation, so the speed advantage compounds across a pipeline rather than being isolated to one call. ## Is Mercury Decide worth using? For teams building agents that need frequent, low-latency judgments, the case is straightforward: it’s cheaper and faster than routing every check through a general-purpose judge model, and independent evaluations suggest it doesn’t sacrifice accuracy to get there. The LegalBench and claim-verification results suggest it holds up on domain-specific judgment tasks, not just toy benchmarks. The caveats are the usual ones for a new model category. Pricing and availability may shift once promotional periods end (it’s currently accessible free on OpenRouter), and like any judge model, its reliability depends heavily on how well the typed questions and state are structured by whoever is building the agent. It’s also not a replacement for a reasoning model. It’s designed to complement one, not substitute for it. ## Frequently Asked Questions ### What makes Mercury Decide different from a regular LLM call? ## Seven tools to build an app. Or just Remy. Editor, preview, AI agents, deploy — all in one tab. Nothing to install. It skips free-text generation entirely. You submit a state and a typed question, and it returns a structured answer with a confidence score in one pass, rather than generating prose you’d then have to parse. ### Can Mercury Decide process images? Yes. Its latest release added vision support, so the “state” passed into the model can include images alongside or instead of text, enabling checks like confirming a screenshot shows the expected UI state. ### How fast is Mercury Decide compared to other judge models? Independent testing from Vals AI measured a median of about 0.16 seconds for a single judgment and about 0.4 seconds for 64 parallel judgments against the same context, making it well suited to batched checks. ### Is Mercury Decide based on the same technology as Mercury 3? Yes. Both are diffusion-based models from Inception Labs. Mercury 3 handles general reasoning and generation, while Mercury Decide is purpose-built for fast structured judgments, and the two are designed to work together in agent pipelines. ### Why does hallucination rate matter for a decision model? A judge model that guesses confidently when it’s wrong can silently break an automated pipeline. A model that signals low confidence instead allows the system to escalate to a human or a stronger model, which is why hallucination resistance is tracked as a core metric for this category.