Moonshot AI now offers multiple powerful models for developers, but Kimi K2.8 Preview and Kimi K3 are not designed for exactly the same purpose.
Kimi K3 is Moonshot AI’s flagship model. It has 2.8 trillion parameters, native visual understanding, up to a 1-million-token context window, and is designed for demanding coding, reasoning, knowledge work, and long-running agent tasks. (Kimi)
Kimi K2.8 Preview takes a different approach. Moonshot describes its overall performance as close to K3 while offering more efficient thinking and strong performance for code completion and routine software development. It also supports up to 1 million tokens of context. (Kimi)
Therefore, the difference is not simply:
K3 is newer, so K2.8 is outdated.
Instead, both models have different positions inside the Kimi ecosystem.
This guide explains the differences that actually matter, including coding, reasoning, context, multimodal support, agent workflows, model availability, and which workloads each model targets.
What Is Kimi K2.8 Preview?
Kimi K2.8 Preview is a coding-focused model currently available through Kimi Code.
Moonshot fully rolled it out on September 11, 2026. Existing users of the kimi-for-coding model ID received the upgrade without needing to change their configuration. (Kimi)
According to Moonshot, K2.8 improves both coding and agent capabilities compared with K2.7 Code. More importantly, it provides significantly more efficient thinking while delivering overall performance close to K3. (Kimi)
Its main strengths include:
- Code completion
- Routine software development
- Coding-agent workflows
- Long-context coding
- Multi-file development tasks
- Efficient reasoning
K2.8 Preview supports low, high, and max reasoning-effort settings. By default, Kimi Code currently uses max for K2.8. (Kimi)
In addition, the model supports image and video input in Kimi Code.
What Is Kimi K3?
Kimi K3 is Moonshot AI’s current flagship model.
Released in July 2026, K3 contains 2.8 trillion parameters and uses architectural techniques including Kimi Delta Attention (KDA) and Attention Residuals. (Kimi)
Moonshot positions K3 for much broader and more demanding workloads, including:
- Long-horizon programming
- Large codebase analysis
- Complex reasoning
- Multi-file refactoring
- Knowledge work
- AI agents
- Visual tasks
- Long-document processing
K3 also has native visual understanding and can support up to a 1-million-token context window, depending on the Kimi Code membership tier being used. (Kimi)
Consequently, K3 is better understood as Moonshot’s highest-capability general model rather than simply another coding model.
Kimi K2.8 vs Kimi K3: Quick Comparison
Here are the most important differences:
| Feature | Kimi K2.8 Preview | Kimi K3 |
|---|---|---|
| Position | Coding-focused model | Flagship model |
| Model ID in Kimi Code | kimi-for-coding | k3 |
| Parameters | Not publicly specified in current Kimi Code docs | 2.8 trillion |
| Maximum context | Up to 1M tokens | Up to 1M tokens |
| Reasoning levels | Low / High / Max | Low / High / Max |
| Default reasoning in Kimi Code | Max | High |
| Coding | Strong | Strongest flagship positioning |
| Routine development | Particularly suitable | Capable |
| Complex reasoning | Strong | Flagship capability |
| Large codebases | Supported | Particularly suitable |
| Image input | Yes | Yes |
| Video input | Yes | Yes |
| Open weights | No separate K2.8 release currently documented | K3 is open-weight |
| Main advantage | Efficient thinking with near-K3 overall performance | Maximum overall capability |
These specifications reflect Moonshot’s current Kimi Code documentation as of September 2026. (Kimi)
1. K3 Is the Flagship Model
The biggest difference is positioning.
Moonshot explicitly describes K3 as its most capable flagship coding model in Kimi Code. Meanwhile, K2.8 Preview is described as delivering overall performance close to K3 with more efficient thinking. (Kimi)
That distinction matters.
K3 targets workloads where maximum capability is important, such as large repository analysis, difficult reasoning, long-horizon engineering, and complicated multi-step tasks.
K2.8, on the other hand, focuses more heavily on practical coding efficiency.
For example, consider these two requests.
A routine development task might be:
“Add validation to this registration form and update the related tests.”
A much more demanding task could be:
“Analyze this large distributed codebase, identify the source of intermittent transaction duplication, determine which services are involved, and propose a safe architectural fix.”
Both models can work with code. However, K3 is positioned more strongly for the second type of problem.
2. K2.8 Focuses More on Efficient Thinking
K2.8’s main advantage is not that it exceeds K3’s flagship capability.
Instead, Moonshot emphasizes thinking efficiency.
The company says K2.8 Preview provides performance close to K3 while improving reasoning efficiency compared with K2.7 Code. (Kimi)
That can matter during everyday development.
Not every coding task requires extremely deep reasoning. Developers frequently ask AI to complete relatively ordinary work, such as adding validation, writing tests, explaining code, updating an API integration, or refactoring a small component.
In those situations, efficiently reaching a useful answer can be more valuable than maximizing reasoning effort.
Therefore, K2.8 is particularly interesting as a practical everyday coding model.
3. Both Support Adjustable Reasoning
K2.8 and K3 both support three reasoning-effort levels in Kimi Code:
Low
High
Max
However, their current defaults differ.
K3 defaults to high, while K2.8 Preview defaults to max. (Kimi)
Reasoning effort lets developers adjust how much thinking the model applies to a task.
For example, low effort may be enough for a straightforward code modification. In contrast, a difficult debugging problem may benefit from high or max reasoning.
Interestingly, Moonshot also notes that disabling thinking for either K3-series requests or K2.8 Preview routes the request to K2.8 Preview without thinking. (Kimi)
That makes K2.8 an important part of Kimi Code’s broader model-routing strategy.
4. Both Can Support 1 Million Tokens
Context size is another area where the models are surprisingly similar.
K2.8 Preview supports up to:
1,048,576 tokens
K3 can also support:
1,048,576 tokens
in its full context configuration. (Kimi)
A million-token context can be valuable for software engineering because AI coding agents may need to inspect large amounts of information.
For example, a project might contain a mobile application, backend services, database models, admin panel, tests, API documentation, and infrastructure configuration.
A long-context model can potentially keep much more of that project available during a single working session.
However, availability differs.
K2.8’s 1M context is available across its supported Kimi Code membership tiers. For K3, the full 1M context currently requires higher-tier access; lower eligible tiers may receive a smaller context limit. (Kimi)
Therefore, simply saying “both have 1M context” misses an important practical difference.
5. K3 Also Has a 256K Version
Kimi Code provides another model ID:
k3-256k
This is still K3, but it has a 262,144-token context window.
Moonshot says it produces the same results as full K3 when working within the 256K context range. At the same time, full 1M K3 consumes approximately twice as much quota as K3-256K. (Kimi)
That gives developers three useful options:
K2.8 Preview → Efficient coding + up to 1M context
K3-256K → Flagship K3 quality for smaller workloads
K3 1M → Flagship capability + maximum context
For example, editing one or two files may not require a million-token context.
In that situation, K3-256K can make more sense than loading the full K3 context configuration.
6. K3 Has a Much Larger Publicly Documented Architecture
One major difference is model scale.
Moonshot has disclosed that K3 contains 2.8 trillion parameters. It is built using a Mixture-of-Experts architecture, Kimi Delta Attention, Attention Residuals, and other architectural improvements. (Kimi)
Moonshot says K3 activates 16 out of 896 experts within its sparse MoE design. (Kimi)
By comparison, the current Kimi Code documentation does not publish a comparable parameter count for K2.8 Preview.
Therefore, it would be misleading to compare the models based on parameter count when Moonshot has not provided equivalent public specifications for K2.8.
More importantly, parameter count alone does not determine practical model quality.
Training, architecture, data, inference techniques, and post-training also matter.
7. Both Support Images and Video
Multimodal capability is another similarity.
Current Kimi Code documentation lists both full K3 and K2.8 Preview as supporting:
Image input
and
Video input. (Kimi)
This can be useful during software development.
For example, a developer could provide a screenshot showing a broken interface and ask the AI to inspect the relevant source code.
Similarly, video input could help demonstrate an animation problem, navigation issue, or unexpected application behavior.
However, there is one important exception.
The smaller k3-256k configuration supports images but does not support video input. (Kimi)
8. K3 Is Better Positioned for Large Codebase Analysis
Both models can handle long context, but Moonshot specifically recommends K3 for tasks such as:
- Large codebase analysis
- Multi-file refactoring
- Long-document processing
- Complex reasoning
(Kimi)
Imagine a company has a large application with hundreds of source files.
A developer asks:
“Replace the old authentication architecture while preserving existing login, refresh-token, session, and logout behavior.”
The AI may need to understand relationships between frontend code, backend APIs, authentication middleware, databases, tests, and configuration.
This is exactly the type of long-horizon engineering task where K3’s flagship positioning becomes relevant.
K2.8 can still work with large contexts. Nevertheless, Moonshot describes it more specifically as strong for code completion and routine development.
9. K2.8 Can Be Better Suited to Everyday Development
Not every developer spends the entire day performing repository-wide migrations.
Most daily work is much smaller.
For example:
“Add this API endpoint.”
“Fix this validation issue.”
“Write tests for this service.”
“Refactor this widget.”
“Explain why this function is failing.”
“Update this component.”
These tasks still require capable coding.
However, they may not require Moonshot’s strongest flagship reasoning every time.
Because K2.8 combines near-K3 overall performance with more efficient thinking, it is specifically positioned for this kind of everyday development. (Kimi)
10. K3 Targets Broader Knowledge Work
K3 is not limited to programming.
Moonshot also highlights:
Complex reasoning
Knowledge work
Long-document analysis
Agent tasks
Visual understanding
alongside coding. (Kimi)
Therefore, K3 is more clearly positioned as a broad flagship intelligence model.
K2.8 Preview is currently presented mainly through Kimi Code, which gives it a stronger developer-oriented identity.
11. K3 Is Open-Weight
Another important difference concerns model availability.
Moonshot released K3 as an open model and announced plans for its full model weights alongside the K3 release. (Kimi)
K2.8 Preview is different.
It is currently delivered through Kimi Code under:
kimi-for-coding
Moonshot’s current Kimi Code documentation does not describe K2.8 Preview as a separately released open-weight model. (Kimi)
Therefore, developers interested in model weights and open-model experimentation should distinguish K3 from the K2.8 Preview service.
12. Model IDs Are Different From Model Names
This small detail can prevent configuration errors.
For K3, the Kimi Code model ID is:
k3
For the smaller K3 context configuration:
k3-256k
For K2.8 Preview:
kimi-for-coding
Moonshot specifically warns developers to use these model IDs, rather than entering model-version names such as K3 or K2.8 Preview. Using the version name directly can cause the request to fail. (Kimi)
This is particularly relevant when configuring Kimi models inside third-party coding tools.
Which Model Is Better for Coding?
There is no single answer for every coding task.
K2.8 Preview is specifically positioned for code completion and routine development, while still offering performance close to K3. (Kimi)
K3 is positioned as Moonshot’s most capable flagship coding model, especially for difficult reasoning, large codebases, multi-file refactoring, and long-horizon tasks. (Kimi)
Therefore, the practical distinction is straightforward.
For normal development work, K2.8 may provide the capabilities you need without always relying on the flagship model.
When the task becomes unusually complex, broad, or reasoning-heavy, K3 provides Moonshot’s highest-capability option.
Which Model Is Better for AI Agents?
Both support agentic coding workflows.
However, K3 is positioned more strongly for complex, long-running tasks requiring deeper reasoning and broader capability. Moonshot specifically highlights long-horizon programming and difficult agent workflows among K3’s strengths. (Kimi)
K2.8 remains highly relevant because Moonshot says its coding and agent capabilities have improved across the board compared with K2.7 Code. (Kimi)
As a result, the complexity of the agent workflow matters more than simply asking whether one model “supports agents.”
K2.8 vs K3: Practical Use Cases
| Task | Model positioning |
|---|---|
| Code completion | K2.8 is specifically suited |
| Routine feature development | K2.8 is specifically suited |
| Small refactoring | Either can handle it |
| Large repository analysis | K3 is specifically recommended |
| Complex debugging | K3’s stronger reasoning is relevant |
| Multi-file architectural changes | K3 is specifically recommended |
| 1M-context coding | Both can support it |
| Complex knowledge work | K3 is more broadly positioned |
| Image-based coding | Both support images |
| Video-based development context | Full K3 and K2.8 support it |
These descriptions reflect Moonshot’s own current model positioning rather than independent claims that one model will always outperform the other on every task. (Kimi)
Frequently Asked Questions
What is the main difference between Kimi K2.8 and Kimi K3?
K2.8 Preview focuses on efficient coding and routine development while delivering overall performance close to K3. In contrast, K3 is Moonshot AI’s flagship model for the most demanding coding, reasoning, knowledge-work, and long-horizon tasks. (Kimi)
Does Kimi K2.8 have a 1M context window?
Yes. K2.8 Preview supports up to 1,048,576 tokens of context in Kimi Code. (Kimi)
Does Kimi K3 have a 1M context window?
Yes. Full K3 supports up to 1,048,576 tokens, although access to the full context size depends on the Kimi Code membership tier. (Kimi)
Which model has more parameters?
Moonshot publicly states that K3 contains 2.8 trillion parameters. The current Kimi Code documentation does not provide a directly comparable parameter count for K2.8 Preview. (Kimi)
Does K2.8 support reasoning?
Yes. K2.8 supports low, high, and max reasoning effort, with max currently used as its default in Kimi Code. (Kimi)
Does K3 support reasoning levels?
Yes. K3 also supports low, high, and max reasoning effort. Its current Kimi Code default is high. (Kimi)
Can K2.8 understand images and videos?
Yes. Kimi Code currently lists K2.8 Preview as supporting both image and video input. (Kimi)
Is Kimi K3 open-weight?
Yes. Moonshot announced K3 as an open model and its official K3 materials describe the release of model weights. (Kimi)
What are the Kimi Code model IDs?
Use k3 for full K3, k3-256k for the smaller K3 context configuration, and kimi-for-coding for K2.8 Preview. (Kimi)
Conclusion
Kimi K2.8 Preview and Kimi K3 are both powerful models, but they occupy different positions in Moonshot AI’s lineup.
K2.8 Preview focuses on practical coding efficiency. It offers performance that Moonshot describes as close to K3, supports up to a 1-million-token context window, provides adjustable reasoning, and is particularly suited to code completion and routine development. (Kimi)
Meanwhile, K3 remains Moonshot AI’s flagship model. Its 2.8-trillion-parameter architecture is designed for demanding coding, complex reasoning, large repositories, long-horizon agent tasks, and broader knowledge work. (Kimi)
Both models support long context, reasoning controls, images, and video in their full Kimi Code configurations. Therefore, the difference is less about whether one model can perform a particular basic task and more about how demanding that task is.
For everyday code completion and routine development, K2.8 is specifically designed around efficient performance. For large codebases, difficult reasoning, extensive refactoring, or long-running agent workflows, K3 provides Moonshot’s flagship capability.
Most importantly, developers should not interpret the model numbers as a simple ranking.
K2.8 is the coding-focused, efficiency-oriented option, while K3 is Moonshot AI’s broader flagship model for its most demanding workloads.
For the latest model specifications and availability, see Kimi Code’s official model documentation.
