Introduction
On September 11, 2026, Moonshot AI quietly launched K2.8 Preview within Kimi Code. This new coding model retains the existing kimi-for-coding model identifier, which means existing clients and third-party developer tools can upgrade automatically without manual configuration changes. Official positioning describes it as a model with comprehensive capability close to K3 and improved thinking efficiency. It supports three tiers of thinking effort: low, high and max. It unlocks a maximum 1 million-token context window alongside image and video input, and this capability is open to all membership tiers.
The release completes a three-tier model stack for Kimi Code: flagship K3, mainstream K2.8 Preview and high-speed K2.7 Code HighSpeed. Any requests submitted with thinking disabled will be routed uniformly to K2.8 Preview. Early side-by-side tests by developers show that K2.8 Preview can finish a single task roughly one-third faster and consume around 30% less cost than K3. However, its one-pass success rate for complex tasks still falls behind K3. Bloomberg reported that Moonshot AI’s annual recurring revenue exceeded $1 billion in August 2026, with a target to hit $2 billion by year-end. K2.8 Preview acts as the core product to carry mainstream traffic while the company scales its commercialization.
1. What is K2.8 Preview: Positioning and Core Specifications
K2.8 Preview serves as the workhorse model within the Kimi Code model family. It sits between the flagship K3 and fast inference variant K2.7 Code HighSpeed. It targets code completion and conventional software development tasks. Based on official Kimi Code documentation dated September 2026, its key specifications are listed in the table below.
| Item | Specification for K2.8 Preview |
|---|---|
| Model ID | kimi-for-coding (reused, zero reconfiguration) |
| Release Date | September 11, 2026, full rollout across Kimi Code |
| Comprehensive Performance | Comparable to K3; thinking efficiency improved notably vs K2.7 Code |
| Thinking Modes | low / high / max three tiers, default max; same tier definition as K3 |
| Context Window | 1048576 tokens (1M), available to all membership tiers |
| Multimodal Input | Image, Video |
| Access Permission | Available to all members |
Official documentation also clarifies a routing rule. After users disable the thinking switch for the K3 series and K2.8 Preview, all incoming requests will be processed by K2.8 Preview in non-thinking mode. This indicates K2.8 Preview is not merely an optional model variant. It functions as the unified base for all non-thinking workloads within Kimi Code.
Developer @notjazii on X platform shared observations on September 11. He stated this new model is smaller and cheaper while delivering near-K3 performance. At that time, Moonshot AI’s official social accounts had not issued formal announcements, even though the model was already live inside Kimi Code.
2. Selection Guide: K2.8 Preview, K3 and K2.7 Code HighSpeed
The core decision rule for developers is straightforward. Use K2.8 Preview for most daily coding jobs. Reserve K3 for highly complex tasks requiring one-pass completion. Choose HighSpeed variants when output latency is the top priority. The table below compares four model IDs from Kimi Code official documentation in September 2026.
| Model ID | Model Version | Description | Context Window | Thinking Level | Access Restrictions |
|---|---|---|---|---|---|
| k3 | K3 | Top flagship reasoning model with 2.8T parameters; high-tier model | 1M | low/high/max, default high | Moderato and above; Allegretto+ unlock full 1M |
| k3-256k | K3 | 256K context variant of K3; consumes roughly half the resources of K3 (1M) | 256K | Same as above | Moderato and above |
| kimi-for-coding | K2.8 Preview | Near-K3 comprehensive capability, higher thinking efficiency | 1M | low/high/max, default max | All members |
| kimi-for-coding-highspeed | K2.7 Code HighSpeed | Identical coding capability to original K2.7; output speed boosted 5–6 times | 256K | Thinking ON | Allegretto and above |
Three practical selection recommendations for developers:
- Entry and mid-tier members:
kimi-for-codingis the only model open to all membership tiers that supports a full 1M context window. After K2.8 Preview launch, the capability ceiling of this tier increases substantially. - High-tier members: K3 (1M) consumes about twice the resources of k3-256k. Developers can handle most tasks using K2.8 Preview and only route the most challenging workloads to K3.
- High-speed mode limitations: The highspeed variant only accelerates model token generation. Latency introduced by tool invocation and script execution remains unchanged. If those steps dominate total runtime, overall speed gains are limited. Also, highspeed version runs at 6x inference speed with triple token consumption.
3. Three Thinking Tiers and Effort Mapping
Like K3, K2.8 Preview supports three thinking levels: low, high and max. Effort values passed from third-party tools map to official defined tiers automatically. The main difference lies in default configuration: K3 defaults to high, while K2.8 Preview defaults to max. The official mapping table is shown below.
Officials recommend keeping the same thinking level within a single conversation session. Switching effort values invalidates cached context. Previously cached prompt content must be reloaded, which increases token consumption. If you have to adjust thinking tiers, start a brand-new conversation session.
4. How to Invoke K2.8 Preview in Third-party Tools such as Claude Code
Calling K2.8 Preview from external tools only takes three simple steps: create an API key, fill the base URL and specify the model ID as kimi-for-coding. Kimi Code API supports both OpenAI and Anthropic compatible protocols simultaneously.
Step-by-step operation guide
- Generate your API Key inside the Kimi Code console.
- Input the corresponding protocol Base URL and API Key in Claude Code, OpenCode, Codex or other compatible developer tools.
- Fill model ID field as
kimi-for-coding. Switch tok3ork3-256kwhen you need K3 capability. - For users of official Kimi Code CLI: type
/modeldirectly to switch models. If the new model is not listed, log out then log back in.
This pattern — keeping the OpenAI-compatible interface and only modifying the model identifier — follows common industry practice. Multi-model unified access platforms use the same pattern. 4sapi provides standardized API endpoints, letting developers switch among mainstream large models without rewriting calling logic, and run side-by-side benchmark evaluation of K2.8 Preview and other models within the same test script.
5. Early Empirical Test: Faster and Cheaper, Yet Lower One-pass Completion Rate Than K3
Developer @notjazii on X platform ran a comparison test on September 11, 2026. He used identical prompts and maximum thinking settings to benchmark the two models, releasing the first batch of public test data.
| Metric | K3 | K2.8 Preview |
|---|---|---|
| Task Completion Status | Completed in one pass | Failed to finish in one pass |
| Total Time Spent | 120 minutes | 80 minutes |
| Estimated Cost | Approximately 14 USD | Approximately 10 USD |
According to this dataset, K2.8 Preview cuts task duration by one-third and reduces total cost by roughly 30%. This aligns with the official claim of improved thinking efficiency. But the developer added supplementary notes in follow-up posts. K2.8 Preview struggled repeatedly to repair issues identified during multi-turn feedback. Its one-pass completion rate and self-repair capability lag behind K3. This sample only reflects one single task and serves as directional reference. Official benchmark scores for K2.8 Preview have not been published.
The practical implication for developers is clear. K2.8 Preview works well as default model for high-frequency coding tasks of moderate complexity. Reserve K3 for complex refactoring or multi-module joint development scenarios requiring one-shot full completion.
6. Competitive Landscape: Moonshot AI Commercialization Timeline After K3 Release
The rollout of K2.8 Preview comes after a wave of commercial growth driven by K3. Bloomberg reported on September 11, 2026 that Moonshot AI disclosed to investors its annual recurring revenue (ARR) surpassed $1 billion in August. The figure stood at $300 million in June, and the company targets $2 billion ARR by year-end. It directly competes against Anthropic and ZAI in this market.
The K3 model released in July fueled this growth. Bloomberg noted K3 achieved top rankings on multiple benchmarks at a cost far below leading US model providers. Developer community enthusiasm also validated this momentum. A project running the 2.8T parameter K3 model on a MacBook Pro with four SSDs went viral on Hacker News on September 8, collecting 277 upvotes. The Next Web reported on August 26 that Moonshot AI aims to capture 30% revenue share from third-party hosted K3 deployments in the US market.
Within this business framework, K2.8 Preview has a defined strategic role. It carries mainstream Kimi Code workloads with near-K3 capability and lower inference cost. K3 remains reserved for high-tier members and high-difficulty tasks. This setup helps Moonshot hit its $2 billion ARR goal while keeping expense ratios controlled.
7. Frequently Asked Questions
Q: What is the difference between Kimi K2.8 Preview and K3?
A: K3 is a flagship model with 2.8T parameters, default thinking level high, accessible for Moderato membership and above. K2.8 Preview delivers comprehensive performance close to K3 and higher thinking efficiency. It defaults to max thinking and is available for all members. Both models support 1M context and image/video input. All requests with thinking disabled will be routed to K2.8 Preview non-thinking pipeline.
Q: Which model does kimi-for-coding correspond to currently?
A: Starting September 11, 2026, kimi-for-coding maps to K2.8 Preview. Previously it pointed to K2.7 Code. Since the model ID remains unchanged, existing clients and third-party tools configured with this ID upgrade automatically, with zero configuration edits required.
Q: Why does token consumption rise after switching to K2.8 Preview?
A: After switching models, old context cache entries are invalid under the new model. Those context segments need to be reloaded, causing higher token consumption immediately after migration. Official advice is starting a new conversation session after switching models, for better performance and lower cost.
Q: What is the maximum context length supported by K2.8 Preview?
A: Up to 1048576 tokens (1M). This 1M window is open to all membership tiers. In contrast, K3’s full 1M context is limited to Allegretto and higher tiers, while Moderato accounts can only access a 256K token limit for K3.
Q: Are there public benchmark scores for K2.8 Preview?
A: As of September 14, 2026, official documentation only states that its comprehensive capability approaches K3. Detailed benchmark metrics have not been released. Available quantitative data comes from independent developer tests with limited sample sizes.
Conclusion
K2.8 Preview achieves a silent capability upgrade by reusing the kimi-for-coding model ID. It unlocks near-K3 capability, a full 1M context window and three adjustable thinking tiers for all members, and becomes the unified processing base for all non-thinking requests within Kimi Code. Early tests show obvious advantages in speed and cost, while complex task one-pass completion still falls short of K3. The model remains in Preview stage, and its capability and default configuration may be adjusted in future updates.
Based on Kimi Code official documents and public press coverage, this article reflects information available up to September 14, 2026. Developers should periodically review official updates for the latest model specifications.
International access: https://4sapi.com
Domestic access: https://4sapi.cn
