Try them all in Kilo. KAT-Coder-Pro V2.5 is free for a week.
AI never sleeps, and that’s especially true of the recent wave of launches from labs that aren’t always in the headlines. Hot on the heels of OpenAI’s three-model GPT-5.6 launch—and the return of Claude Fable—Meta, MoonshotAI and Kat Coder all just released models that are turning heads.
And all of these new models are live wherever you use Kilo: the CLI, JetBrains extension, VS Code extension, Cloud Agents and beyond.
New Kat Coder Tiers: Pro & Air #
Kat Coder released two models this week for V2.5: Pro and Air. We’ve deployed both versions of KwaiPilot’s new architecture so you can route your tasks with precision:
Kat Coder Pro v2.5 (Free this week!): The powerhouse flagship. Built for complex, repository-level engineering challenges. Hand it a GitHub issue, and it will locate the files, implement the logic, and run tests until the issue is solved.Kat Coder Air v2.5: The speed-optimized tier. Air brings the same core agentic planning and beautiful front-end aesthetic generation into a faster, highly cost-effective model for rapid everyday tasks.
Technical Highlights of the v2.5 Architecture #
According to the KwaiPilot team’s newly released technical report on arxiv, building a world-class agentic model requires a complete rethink of training systems:
The AutoBuilder Engine: To train an agent to solve real-world problems, it needs a reliable sandbox. KwaiPilot developed AutoBuilder, an automated build-verification loop that reconstructs real repositories into isolated, executable environments. This pipeline boosted environment-construction success rates from 16.5% to57.2%, yielding over 100,000 verified environments across 12 languages.** Process-Aware Trajectory Filtering:**Instead of training the model purely on whether a test ultimately passed (which can reward lazy shortcuts), Kat Coder was trained using a process-aware pipeline. It scores the quality of the agent’s exploration, file localization, and recovery behavior, turning even near-miss failures into valuable training data.Harness Randomization To ensure the agent is highly adaptable, the training framework features harness randomization. By exposing the model to various workspace layouts, tool protocols, and context formats during reinforcement learning, Kat Coder learns to adapt to any developer environment.
Enter Kimi K3: The 1-Million Token Titan #
As if the Kat Coder launch wasn’t enough to keep developers awake all night, Moonshot AI dropped Kimi K3 early this morning. Live right now on their platform and accessible via Kilo, this massive model is a direct challenge to Western frontier giants.
The headlining breakthrough of Kimi K3 is its colossal 1-million token context window—a massive 4x expansion over the older K2 series. Powered by a highly sophisticated Mixture-of-Experts (MoE) architecture running on an estimated 2 to 3 trillion parameters, Kimi K3 is officially China’s largest AI model to date. This massive context and scale allow developers to feed entire enterprise codebases, hundreds of pages of documentation, or highly complex system logs into a single prompt without losing a single line of detail.
Beyond its massive memory, Kimi K3 introduces a highly anticipated framework: Agent Swarm Technology. This architecture allows the model to spin up and coordinate up to 300 sub-agents in parallel. For complex multi-step tasks like refactoring sprawling microservice environments or carrying out complex cross-dependency updates, these sub-agents plan, delegate, and execute work asynchronously.
Early testers at Kilo and around the web are already calling the K3 drop another “DeepSeek R1 moment,” signaling that the performance gap between domestic models and US frontier labs is closing fast.
Remember Llama? With Muse Spark 1.1, Meta is Back in the Race #
While Moonshot is scaling up parameter sizes, Meta Superintelligence Labs has thrown its own superstar into the ring: Muse Spark 1.1. Succeeding its predecessor in record time, Muse Spark 1.1 represents Meta’s aggressive leap from consumer assistants to high-end, developer-facing APIs with safety and security in mind (read their full report for details). Muse Spark 1.1 is already in the top 10 on KiloBench. This natively multimodal reasoning model boasts a massive 1-million token context window and is built from the ground up to plan, navigate, and execute complex workflows across external apps and services rather than just answering single-turn questions.
The core superpower of Muse Spark 1.1 lies in its ability to orchestrate multi-agent projects and operate computer interfaces directly. Instead of processing a desktop step one click at a time, it is trained to construct a macro plan, write execution scripts on the fly, and spin up parallel sub-agents to handle tasks like web-scraping, database entry, and frontend design simultaneously.
Upgrade Your Engineering Loop Today #
Whether you’re a solo dev or a part of an large enterprise, Kilo has the tools to supercharge your flow. The landscape is converging—large models, and cheaper pricing, seem to be coming from everybody these days.
Context and Size: Muse Spark 1.1, Kimi K3, and frontier models like Claude Opus 4.8 lead the pack with sprawling1-million token context windows to ingest massive codebases and system logs. In contrast, Kat Coder Pro v2.5 stays lightweight, focusing its repository-level magic on a highly efficient256K context window.** Pricing:**Developer economics are shifting rapidly. While premium reasoners like Claude Opus 4.8 ($5.00/$25.00) and Kimi K3 ($3.00/$15.00) represent significant investments, models like Muse Spark 1.1 ($1.25/$4.25) and Kat Coder Pro v2.5 ($0.74/$2.96—free this week****in Kilo!) offer incredibly deep agentic capabilities at a fraction of the cost.
Give Kat Coder Pro v2.5 a try for free today, and see how it works for your agents and subagents in collaboration with these new models from Meta and Moonshot AI—and see how they start to trend on the Kilo Leaderboard.
We can’t wait to hear your thoughts!