AI & Technology
Moonshot AI has unveiled its latest iteration, the Kimi K2.7 Code, an innovative open AI coding model.
This new version boasts a remarkable 21.8% improvement over its predecessor, Kimi K2.6, and outperforms Claude Opus 4.8 in a specific benchmark, as China approaches the forefront of AI development.
Moonshot AI has introduced the Kimi K2.7 Code, a state-of-the-art coding model designed to facilitate the rapid writing of software.
Capable of executing complex coding tasks autonomously, this model demonstrates a striking 21.8% enhancement over its predecessor, Kimi K2.6, on key performance metrics.
This development arrives at a time when Chinese enterprises are closing the gap with leading US research facilities in the realm of open coding models.
On June 12, 2026, Moonshot AI announced that the new model is accessible through three platforms: Hugging Face, the Kimi API, and the Kimi Code platform. An API, or Application Programming Interface, enables software to communicate with the model effortlessly over the internet.
Details of the Announcement
The Kimi K2.7 Code is classified as an “agentic” coding model, referring to its capability to strategize and execute extensive tasks in a sequential manner.
Unlike previous iterations that answered queries in isolation, Moonshot AI developed this model specifically for comprehensive, long-term projects requiring multiple stages for completion.
The model features “open weights,” allowing for the public availability of its core files. Under a Modified MIT license, anyone can download and deploy it independently, using tools such as vLLM, SGLang, or KTransformers for local execution.
Constructing the Model
The Kimi K2.7 Code utilizes a Mixture-of-Experts (MoE) architecture. In this configuration, the AI comprises numerous specialized “expert” modules, activating a select few for any given task, thereby optimizing computational efficiency.
The model encompasses 1 trillion total parameters, which are the internal settings the model refines during training.
A greater number of parameters generally indicates a more robust model; however, only 32 billion parameters are activated for each token processed.
A token represents a piece of text, such as a word or partial word, enabling efficient performance while maintaining lower operational costs.
It includes 384 experts in total. For processing each token, it activates 8 experts and 1 shared expert. The model comprises 61 layers and incorporates a vision encoder, dubbed MoonViT, which adds an additional 400 million parameters.
This encoder empowers the model to interpret images and videos, alongside textual data. Additionally, it provides a context window that can accommodate 256K tokens, equal to the content of a small book.
Benchmark Performance
Benchmarks serve as standard evaluations to gauge the effectiveness of AI models. Moonshot AI released comprehensive benchmarking results, contrasting the performance of Kimi K2.7 Code with that of Kimi K2.6.
In Kimi Code Bench v2, the new model achieved a score of 62.0 compared to the prior version’s 50.9, marking the aforementioned 21.8% enhancement.
The K2.7 performed admirably across various assessments. For instance, on the MCP Mark Verified benchmark, it outpaced Claude Opus 4.8 with a score of 81.1 versus 76.4 for the latter.
Moreover, Moonshot AI noted that the new model consumes approximately 30% fewer “reasoning tokens” than its predecessor. This reduction in hidden computational steps results in quicker responses and lower costs.
Benchmark Results & Specifications
| Benchmark (as reported) | K2.7 Code | K2.6 | Change |
|---|---|---|---|
| Kimi Code Bench v2 | 62.0 | 50.9 | +21.8% |
| Program Bench | 53.6 | 48.3 | +11.0% |
| MLS Bench Lite | 35.1 | 26.7 | +31.5% |
| Kimi Claw 24/7 Bench | 46.9 | 42.9 | +9.3% |
| MCP Atlas | 76.0 | 69.4 | +9.5% |
| MCP Mark Verified | 81.1 | 72.8 | +11.4% |
Scores as reported by Moonshot AI; on the MCP Mark Verified benchmark, K2.7 Code (81.1) surpasses Claude Opus 4.8 (76.4).
| Spec | Detail (as reported) |
|---|---|
| Type | Mixture-of-Experts (MoE) coding model |
| Total parameters | 1 trillion |
| Active per token | 32 billion |
| Experts | 384 total; 8 + 1 shared per token |
| Layers | 61 (including 1 dense layer) |
| Vision encoder | MoonViT (+400M params); text, image, video |
| Context window | 256K tokens |
| Quantization | Native INT4 |
| License | Open weights, Modified MIT |
| API price | $0.95 / M cached input; $4.00 / M output tokens |
| Download size | ~595 GB (Hugging Face) |
In essence, a large, openly accessible coding model that utilizes only a fraction of its capacity per token, rendering it economically viable. Kimi K2.7 Code has shown improvements over K2.6 in all benchmarks reported by Moonshot AI. Source: Moonshot AI.
It is crucial to note that the model consistently operates in “thinking mode,” with no option to deactivate this feature.
Furthermore, it employs fixed sampling parameters, specifically a temperature of 1.0 and a top_p of 0.95. By default, the model can return up to 32,768 tokens in one response.
Comparative Analysis with Competitors
Moonshot AI has refrained from claiming outright dominance of its new model. It acknowledges that GPT-5.5 continues to outperform K2.7 Code in most evaluated scenarios.
However, against Claude Opus 4.8, K2.7 achieved victory in the MCP Mark Verified benchmark. This mixed outcome illustrates the competitive nature of open models from China as they contend with the most prominent closed models from American research facilities.
What is the Kimi K2.7 Code?
Kimi K2.7 Code is an advanced AI coding model developed by Moonshot AI, specifically fashioned to facilitate software development and manage extensive, multi-step coding tasks autonomously.
Is the model free to use?
The model’s files are publicly available under a Modified MIT license, enabling users to download and execute them independently at no cost. Usage of the official API, however, incurs charges, priced at $0.95 per million cached input tokens and $4.00 per million output tokens.
How does it compare to Kimi K2.6?
In the Kimi Code Bench v2 test, Kimi K2.7 Code secured a score of 62.0, while K2.6 attained a score of 50.9, resulting in a substantial enhancement of 21.8%. The new model outperformed its predecessor across all reported assessments.
Can it outperform GPT-5.5 or Claude?
Not in every instance. According to Moonshot AI, GPT-5.5 consistently scores higher across the majority of benchmarks. However, Kimi K2.7 Code did surpass Claude Opus 4.8 in the MCP Mark Verified assessment.
Significance for India and Founders
Open coding models, such as Kimi K2.7 Code, hold substantial significance for Indian entrepreneurs and developers.
The availability of model files allows teams to download and operate the model on their own systems, consequently reducing operational costs and maintaining confidentiality for proprietary code.
Additionally, the emergence of robust open models offers increased flexibility. Startups are no longer limited to a few US providers, as affordable, capable coding AI enables smaller teams to accomplish more with fewer resources.
Nonetheless, the requirement for powerful hardware to run the model remains a consideration, with the download size approaching 595 GB. As a result, many small teams may initially resort to the API for access.
The overarching trend is unmistakable: the rapid advancement of China’s open models continues to bring them closer to the best-in-class closed systems, enhancing the toolset available for developers globally.
Ultimately, while Kimi K2.7 Code may not dominate every assessment, it stands as a formidable, open, and cost-effective alternative in the evolving landscape of AI coding solutions, advancing the capabilities of China’s open models.
For founders vigilant about budget considerations, the rise of genuine competition within the coding AI sector is undeniably promising.
Source link: Voice.lapaas.com.
