Cast AI integrates MiniMax M3 into Kimchi Coding agent

Cast AI has added MiniMax M3 to its Kimchi Coding autonomous coding agent. Kimchi is the first coding agent to offer access to the model.

The move makes MiniMax M3 the default builder model in Kimchi's orchestration layer, which assigns coding tasks across different AI models based on complexity and cost. The product is being rolled out in phases through an early access programme.

The announcement reflects a wider shift in how businesses deploy AI models for software development and related work. Rather than relying on one model for every job, companies are increasingly directing tasks to a mix of open-weight and commercial systems, using the more expensive options only where they are most needed.

Kimchi is an open-source, terminal-native coding agent and AI assistant for software development workflows. It can run as a serverless deployment on Cast AI infrastructure or inside a customer's own environment across AWS, GCP, Azure or on-premise systems, including air-gapped setups.

Kimchi includes an orchestration engine that routes each task to the best-fit model. The system also scores generated code and uses continuous feedback loops to check output while managing token consumption. Customers can set spending limits for individual API keys or across organisations, with automatic termination of runaway agent loops and a dashboard to track token use by user, team and project.

Model economics

MiniMax M3 is positioned as an open-weight model for long-context coding work. According to Cast AI, it scored 59% on SWE-bench Pro, a benchmark based on real GitHub issues, and outperformed several commercial rivals on that measure.

Cast AI also said MiniMax M3 uses a sparse attention architecture that cuts per-token compute at a one-million-token context to one twentieth of previous levels while increasing decoding speed 15-fold. For organisations running longer coding tasks or traversing large codebases, that would affect both cost and completion time.

In continuous shadow-mode evaluations against a baseline using only commercial models, Cast AI said Kimchi delivered costs 2.5 times lower while matching or exceeding quality on specification match and test pass rates.

Laurent Gil, President and Co-Founder of Cast AI, linked the launch to the cost of software development tools built on large language models. "We built Kimchi to give every developer frontier-quality AI coding without frontier-sized bills or data risk, thanks to our built-in token-optimized model orchestrator," he said.

He added: "MiniMax M3 is the model Kimchi was designed for: open weights, frontier performance, and economics that work at scale."

Enterprise use

The addition of MiniMax M3 also points to a growing enterprise focus on where models run and who controls the underlying data. Cast AI said Kimchi can be deployed in a customer's own cloud or on-premise environment, and supports air-gapped installations for organisations with tighter security requirements.

The offering is prepared for compliance requirements including SOC 2, GDPR, ISO 27001 and PCI-DSS. That positioning is likely to appeal to larger businesses assessing whether autonomous coding tools can be used in internal engineering systems without sending sensitive code and data to external providers.

MiniMax framed the partnership as a route into enterprise development teams looking for alternatives to commercial-only model stacks. "Cast AI's Kimchi Coding is an ideal partner to bring M3 to teams that demand both performance and sovereignty," said Leanna Ren, Vice President of Global Marketing at MiniMax.

Ren added: "M3 was built to push the boundaries of open-weight models. Kimchi's model orchestrator and data sovereignty controls make it the natural home for M3 in the enterprise. Together, we're giving developer teams frontier-quality AI coding inside their own environment, at a fraction of the cost of commercial models alone."

Cast AI, which said it reached a valuation of more than USD $1 billion after an investment from Pacific Alliance Ventures, sells automation software for cloud-native and AI infrastructure. Its customers include BMW, Cisco, FICO, HuggingFace and Swisscom.

For Kimchi users, the immediate change is that MiniMax M3 now sits at the centre of the coding agent's task-routing system, giving customers another option as companies weigh cost, model choice and data control in AI-assisted software engineering.