Spectre AI: Unveiling the New Free Stealth Model Shaking Up the Coding World

A new mysterious AI model, Spectre, has quietly emerged, and it’s making waves in the AI coding community. Announced as a “free stealth model from a top AI model lab,” Spectre is available exclusively through the Kilo Code AI agent and boasts some impressive capabilities. This article delves into what Spectre is, how it performs, and how you can start using it for your software development tasks today.

[00:08.536]

[Kilo Code Blog post announcing Spectre AI model]

The AI world is buzzing with the release of a new model named Spectre. Described as a stealth model from one of the top 10 AI model labs, it is currently available for early access through Kilo Code, a prominent AI coding agent. This free, stealth model offers unlimited usage for a limited time, making it an exciting new tool for developers to explore.

[00:21.826]

[Model details of Spectre AI from Kilo Code’s blog]

Spectre is specifically optimized for software development tasks and comes with some powerful specifications. It features a massive 256,000-token context window, allowing it to process large amounts of code and information at once. The model has a maximum output of 8,000 tokens per step, which is substantial for generating complex code blocks. While it currently lacks image input or multimodal support, the creators have indicated that updates are expected soon.

This might be your new favorite AI model.

[00:51.526]

[Kilo Code AI Agent extension in VS Code marketplace]

To start using Spectre, you need to install the Kilo Code AI Agent extension. You can find it by searching for “Kilo Code” in the VS Code marketplace. Once installed, the Kilo Code icon will appear in your sidebar, giving you access to the agent’s features.

[01:05.126]

[Configuring Spectre model in Kilo Code AI Agent settings]

Enabling the Spectre model requires a few simple steps within the Kilo Code settings. First, navigate to the settings and create a new configuration profile. Name it whatever you like, for example, “Spectre”. Next, set the API Provider to Kilo Gateway. Finally, in the model selection dropdown, search for and select spectre. After saving these changes and making the profile active, you’ll be ready to use the model.

[01:14.376]

[Making the new Spectre configuration profile active in Kilo Code]

An interesting characteristic of Spectre is that it appears to be a non-reasoning model. During operation, it doesn’t display any intermediate “thinking” steps or reasoning traces, which are common in other advanced models. It simply proceeds directly to the execution of tasks. Currently, the model is offered with unlimited usage and no rate limits, providing an excellent opportunity for extensive testing and use.

[01:38.936]

[Spectre model running a task in Kilo Code within VS Code]

In terms of performance, Spectre is surprisingly effective. One of its standout features is its seamless and accurate tool-calling capability. The model is also very concise; it is not verbose and avoids unnecessary explanations. Instead of stating its intention, like “I’m going to edit a file,” it just performs the action directly. This direct, action-oriented behavior is efficient and similar to what has been observed in other high-performing models like MiniMax.

[02:13.916]

[Task prompt for solving Advent of Code question with Spectre]

To put Spectre to the test, it was tasked with solving the Advent of Code 2025 Day 1 problem. The model handled the first part of the puzzle seamlessly. Although it initially got a bit confused on the more complex second part, it impressively corrected itself autonomously and delivered the correct solution. This self-correction capability highlights its robustness. In comparison, other powerful models like GLM-4.6 and MiniMax failed on the second part of the same task, while Kimi succeeded, placing Spectre in an elite group of capable coding models.

[03:22.386]

[Mistral AI’s blog post about their non-production license]

The strong performance and stealthy release have led to speculation about Spectre’s true identity. The evidence points towards it being a new model from Mistral AI, possibly Mistral Large or Codestral. This theory is supported by the model’s behavior, which aligns with a new generation of models that are highly proficient in tool use, a significant improvement over Mistral’s previous releases. Mistral has recently shown a renewed commitment to open-source by moving back towards Apache licenses, and a powerful free model like Spectre would mark a strong comeback.

[05:40.896]

[Kilo Code blog post highlighting key features of Spectre]

On agentic benchmarks, Spectre’s performance is comparable to GLM, which is impressive for a non-reasoning model. It shows almost no tool call errors or diff edit failures. However, it does have its limitations. The model struggles with Godot game engine development and general frontend tasks. Conversely, it seems to be exceptionally skilled at backend development, writing simple, effective, and non-overcomplicated code.

This is raw model performance that doesn’t use chain of thought or reasoning steps.

[06:56.576]

[Kilo Code blog’s “Our Impressions So Far” section on Spectre]

It’s crucial to remember that Spectre’s impressive results are achieved without explicit reasoning steps. Unlike models such as MiniMax, GLM, Kimi, and DeepSeek, which now incorporate reasoning to enhance performance (often at the cost of speed and latency), Spectre delivers raw, direct performance. The future promise of multimodal capabilities and a potential reasoning variant makes Spectre a model to watch closely as the open-source AI landscape continues to evolve.