Mercury Edit 2
Mercury Edit 2: The Fastest Diffusion-Based LLM for Next-Edit Prediction and Real-Time Coding
Mercury Edit 2 is a purpose-built diffusion Large Language Model (dLLM) designed for the most latency-sensitive coding workflows. By generating tokens in parallel, Mercury Edit 2 provides lightning-fast next-edit predictions that feel instantaneous. This advanced model upgrades previous iterations to deliver higher quality, more targeted code suggestions with a 48% higher acceptance rate. Optimized for the Inception Platform and integrated with editors like Zed, Mercury Edit 2 offers superior speed and quality across multiple programming languages and development scenarios.
2026-04-06
--K
Mercury Edit 2 Product Information
Introducing Mercury Edit 2: The Fastest Reasoning LLM for Next-Edit Prediction
In the rapidly evolving world of software development, latency is the enemy of flow. Today, we are proud to introduce Mercury Edit 2, a purpose-built diffusion LLM (dLLM) specifically engineered for the most latency-sensitive component of modern development workflows: next-edit prediction.
Mercury Edit 2 represents a significant upgrade to our previous next-edit models. It is designed to complement our existing auto-complete endpoints by providing a predictive layer that anticipates your next move within the codebase. By utilizing cutting-edge diffusion technology, Mercury Edit 2 generates tokens in parallel, ensuring that predictions arrive fast enough to feel like a natural extension of your own thought process.
What is Mercury Edit 2?
Mercury Edit 2 is a specialized Large Language Model that focuses on predicting the very next change a developer will make. Unlike traditional models that suggest code one token at a time, this diffusion LLM allows for high-speed parallel generation.
Using your recent edits and comprehensive codebase context, Mercury Edit 2 analyzes the state of your project to suggest the most logical next step. Whether you are refactoring, renaming variables, or implementing a new feature, you can simply press Tab to accept the suggestion and keep your momentum.
Key Features of Mercury Edit 2
Diffusion-Based Token Generation
The standout feature of Mercury Edit 2 is its use of diffusion to generate tokens in parallel. This architectural choice drastically reduces latency, making it the fastest reasoning LLM for edit predictions on the market.
High-Quality Training and Alignment
To ensure the utility of Mercury Edit 2, we utilized a sophisticated training pipeline:
- Curated Datasets: The model was trained on high-quality edits across a broad range of programming languages.
- Human Preference Alignment (KTO): We leveraged an unpaired reinforcement learning method called KTO to align the model with actual human preferences.
- Reduced Distraction: This alignment makes Mercury Edit 2 27% more selective, ensuring suggestions are targeted and less likely to distract the developer.
Superior Quality and Performance
According to our internal and open-source benchmarks (including Instinct, FIM, and NEP), Mercury Edit 2 delivers:
- 48% higher acceptance rate compared to previous models.
- Superior speed when compared to custom next-edit models and speed-optimized frontier models.
- High accuracy across tasks like variable renaming, refactoring, and line completion.
Use Cases for Mercury Edit 2
Mercury Edit 2 is versatile and adapts to various coding scenarios including:
- Next-Edit Prediction: Automatically anticipating the next block of code or modification based on your recent history.
- Refactoring: Streamlining the process of restructuring existing code with high-confidence suggestions.
- Variable Renaming: Quickly updating identifiers across a contextually aware scope.
- Feature Implementation: Accelerating the creation of new functionality by predicting the logical flow of new code.
- Line Completion: Providing instant completion for repetitive or predictable code patterns.
How to Use Mercury Edit 2
Getting started with Mercury Edit 2 is straightforward, whether you are an individual developer or an enterprise team.
- Access via Inception Platform: Mercury Edit 2 is available now on the Inception API Platform.
- Integrate with Zed: Users of the Zed code editor can configure the model to be their primary edit prediction provider.
- Use the API: Developers can integrate Mercury Edit 2 into their own tools using our API. The pricing is highly competitive:
- Input Tokens: $0.25 / 1M tokens
- Output Tokens: $0.75 / 1M tokens
- Cached Input: $0.025 / 1M tokens
- Claim Free Tokens: Every new account on the Inception API Platform is automatically granted 10 million FREE tokens.
Pro Tip for Zed Users: Unlock one free month of Mercury Edit 2 suggestions by using the API key:
sk_ae471146ea60fc117c131b574b00ba96.
FAQ
Q: What makes Mercury Edit 2 different from standard auto-complete? A: While auto-complete suggests the next few characters or words, Mercury Edit 2 is a next-edit prediction model. It uses the context of your recent edits and the entire codebase to predict the next meaningful change you will make, often spanning multiple lines or specific logic shifts.
Q: Why does Mercury Edit 2 use diffusion? A: Diffusion allows the model to generate tokens in parallel rather than sequentially. This is what makes Mercury Edit 2 exceptionally fast, providing the low latency required for a seamless coding experience.
Q: How was the model's accuracy measured? A: We used a suite of four benchmarks: Instinct, Fill-in-the-middle (FIM), Next-edit Prediction (NEP), and an internal benchmark. These tests use LLM-as-a-judge and functional test cases to ensure the suggestions match human-written gold standards.
Q: Is there a free trial available? A: Yes! New accounts on the Inception API Platform receive 10 million free tokens, and Zed users can access a free month using the specific promotional API key provided above.
Q: How can I provide feedback or get support?
A: You can reach out to the team at [email protected] or join our Discord community for early access and support.
