L'ère du seul agent est officiellement terminée : si vous ne pouvez pas battre un, passez à 300

Agent finally bids farewell to “lone wolf” and enters a second stage of evolution?

Just this morning, the Dark Side of the Moon officially released and open-sourced the latest flagship model of the Kimi series — Kimi K2.6, less than three months after the previous version K2.5 was launched. After release, it generated tremendous buzz, with the official account’s viewership reaching 4 million.

Currently, agents often struggle when handling complex engineering projects; although they excel at completing specific tasks independently, team collaboration still needs improvement. How to break through this limitation has become the core goal of Kimi K2.6.

The new version explores how to stimulate agents’ teamwork capabilities: further strengthening the Agent Swarm (agent cluster) feature introduced in K2.5, by adapting frameworks like OpenClaw to enhance proactive work. The brand-new Claw Group (Claw 群组) adds organizational collaboration abilities. This systematic stacking of capabilities constructs an AI system closer to human teams.

To achieve all this, the underlying model must be sufficiently powerful. Kimi K2.6 shows clear progress in core abilities such as general agents, coding, and image understanding. Tests like Humanity’s Last Exam, SWE-Bench Pro (closer to real development scenarios), and DeepSearchQA (assessing deep retrieval capabilities) all demonstrate that K2.6 firmly leads its competitors.

Even when comparing K2.6 with closed-source models like GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro, it holds its ground, and some metrics even surpass them.

Artificial Analysis, a large model evaluation platform, announced the latest results: “Kimi K2.6 becomes the new king among open-source models”!

After launching Kimi K2.6, the large model aggregation platform OpenRouter gave high praise, stating that the new Dark Side of the Moon model emphasizes long-sequence programming capabilities, designed specifically for agent scenarios requiring continuous execution. Compared to traditional chatbots, it functions more like a “systems engineer,” capable of breaking down complex tasks, executing step-by-step, and continuously optimizing during the process.

Some netizens exclaimed that this flagship Kimi model is ridiculously powerful; it can already compete with GPT-5.4 in coding, costs much less than Opus 4.7, and is open-source and free to use. Now, every few months, a new open-source model approaches the performance of closed-source GPT and Claude. “It feels like open-source models are really catching up, and China is leading the pace.”

12 hours nonstop, 300 agents working simultaneously

Is the ultimate form of Agent here?

This time, Kimi K2.6 continues to push forward in the coding domain. A few days ago, overseas discussions heated up over the low-profile launch of Kimi K2.6-Code-Preview, with high expectations for the official release of K2.6.

As the most powerful model in the Kimi series for coding so far, Kimi K2.6 has achieved breakthroughs in long-range coding capabilities, helping to push software development automation into deeper engineering stages.

For example, Kimi K2.6 can successfully download Qwen3.5-0.8B locally on Mac and run it. It does not follow common tech stacks but rewrites reasoning processes in the niche Zig language and continuously optimizes, demonstrating the model’s generalization ability.

The entire process lasted over 12 hours, during which it called tools more than 4,000 times, iterated 14 times. Through constant tuning and restructuring, inference speed jumped from about 15 tokens/sec initially to approximately 193 tokens/sec, ultimately about 20% faster than local large-model chat applications like LM Studio.

The focus of this Kimi K2.6 upgrade is to further enhance the agent cluster’s collaborative output ability. Simply put, this feature aims to clarify “how agents work together.”

What can it do now? Kimi K2.6 can automatically decompose a complex task, assign different specialized agents to handle search, in-depth research, document analysis, long-form writing, etc., then stitch the results together to continue progressing.

Under this mechanism, a single run can complete the entire chain: from raw data and web content to PPTs and spreadsheets, all automatically generated, without back-and-forth tool switching or manual relay.

Meanwhile, the underlying architecture of the agent cluster has been expanded to support up to 300 sub-agents simultaneously, completing 4,000 steps of collaboration, with parallel capabilities reaching a new level. As the scale increases, the role of AI also shifts: it begins to take over the entire process and directly produce systematic results.

The agent cluster deconstructed and reused a high-density astrophysics paper with visual data, generating about 7,000 words of research report, 20,000 data points, and 14 charts.

To evolve AI into a 24/7, uninterrupted cyber employee without human intervention, Kimi K2.6 has made deeper adaptations to frameworks like OpenClaw and Hermes Agent.

For this, Kimi K2.6 further squeezes the model’s autonomous execution capabilities: whether it’s the accuracy of API calls, stability during long-term operation, or safety protections during complex research tasks, K2.6 performs notably.

In Vibe Coding, Kimi K2.6’s website design is more eye-catching. The generated site, especially the first screen, has a strong visual impact, with consistent style. The addition of interactive elements and scrolling effects also helps attract users to stay longer.

Besides front-end design, Kimi K2.6 also brings surprises to back-end developers: it now supports Kimi account login and form data collection. You can create event registration pages and easily view backend registration info. This makes front-back integration smoother.

Currently, Kimi K2.6 is the default model for Kimi web version, app, and Kimi Code programming assistant. Go ahead and try it.

Hands-on testing, impressing everyone

Without further ado, let’s directly test some cases and see how they perform.

The first part of testing uses “K2.6 Agent,” evaluating from both practicality and aesthetics, to see if it can produce eye-catching front-end effects.

Does anyone like “Persona 5”?

This is a highly recognizable artistic style, a visual violence aesthetic cloaked in manga. It challenges aesthetic inertia with highly irregular design, embedding themes of “rebelling against social mediocrity” directly into pixels and lines. It perfectly blends 2D design with 3D space, deep integrating manga symbols and visual expression.

If we opened a P5-style small bar, what would its homepage look like?

We found that during front-end webpage construction, Kimi K2.6’s intelligence performs thorough testing, even simulating clicks:

Additionally, we made a little Easter egg: Kimi K2.6 referenced the opening video of “Persona 5 Royal,” creating a short animation without any material provided.

We continued to demand a different style: “Design a visually impactful homepage for an e-commerce platform, with a top navigation bar including brand logo, search box, shopping cart, login/register buttons; a main banner (Hero Section) showcasing major promotions, bestsellers, or seasonal discounts; below the Hero Section, display recommended products or categories; and at the bottom or a prominent area, show user reviews of selected items.”

A single generation produced a homepage with very high completeness. Though slightly imperfect, we believe small issues can be fixed in one iteration, and the overall quality remains excellent.

Next, we tested the agent cluster’s ability by creating a promotional brochure for Stanford University’s “2026 Artificial Intelligence Index Report,” requiring it to deliver web pages, tables, and PPTs without any additional info or documents, testing the cluster’s collaborative writing performance.

We noticed each agent had its own badge, role description, and profile. When using the agent cluster, you really feel like a strategic CEO mobilizing all resources, assigning tasks to the right people, forming a working team that automatically executes tasks. It’s almost like having “reliability” written on their badges.

It ultimately delivered all the content we needed: a shiny webpage, efficiently formatted PPT, and serious data tables.

Is multi-agent collaboration the future?

These series of tests demonstrate the powerful strength of Kimi K2.6 as the “base model” of the Agent era.

In the current “lobster craze” sparked by OpenClaw, the newly launched Claw Group points to a clear path for next-stage agent evolution.

Currently, Claw Group is in limited internal testing.

This feature marks a new era of agent collaboration. You can connect various agents running locally, on mobile, or in the cloud, each equipped with tools, skills, and memory, working together in a “group” to advance tasks.

Here, K2.6 acts more like a dispatcher: who is good at retrieval, who handles analysis, who produces content, it divides tasks based on capabilities. If any link gets stuck, it can detect and reassign tasks or switch personnel to keep the process moving.

Imagine preparing a complex report or developing a multi-layered project: the Claw group’s agents will discuss, coordinate, and adjust like a team of professionals, ultimately delivering a precise, comprehensive result.

This innovation not only breaks through traditional individual agent execution modes but also pushes organizational intelligence forward. Its emergence makes “multiple AI agents working together” more of a reality.