Agentic AI

The training pipeline involves generating nine specialized domain experts (3 reasoning levels: low, high, max) for three general domains (general tasks, deep research, coding). These are combined using multi-tier on-policy distillation into a single student model.

A novel RL technique where the system samples $k$ rollouts from $n$ prompts. It uses an additional parameter $\lambda$ to measure completed rollouts within a budget, updating weights on that subset and pushing incomplete rollouts into subsequent iterations.

The Kimi K3 model utilizes an 'extensible token markup language' for its chat template. This design aims to be easy to learn during SFT while remaining flexible enough to accommodate future modalities without re-engineering the core template.

The system employs sophisticated resource management, including a dynamic rollout auto throttling scheduler based on KV cache pressure and a method for collocating RL training and inference by automatically balancing GPU resources between the two tasks.

The goal is to sell the final business outcome (e.g., higher throughput of sales) rather than selling a piece of software or the time of an engineer. This model allows companies to land large contracts that self-serve motions cannot reach.

An FDE function is only necessary when selling something very technical (like an app building platform) to a non-technical buyer. If the product is simple or the buyer is highly technical, other GTM strategies may suffice.

To scale FDE successfully, engineers must build on a platform of shared primitives. If every engineer builds entirely from scratch for each customer, the function becomes an unmaintainable 'dev shop,' leading to massive maintenance costs.

The current shift toward agentic and customizable platforms means that nearly all companies may face the situation of selling complex solutions to non-technical customers, making FDE a more common motion.

Synthetic data is used not to replace organic data but to complement it by extracting and projecting implicitly hidden features (e.g., implicit rationale, planning, structure). Pipelines are modular, allowing complex tasks to be broken down into stages like multi-stage workflows, cross-domain porting (e.g., math problems to code), or multi-turn chats.

To ensure correctness at scale, the team implemented model replica hashing checks; if hashes over weights do not match across distributed replicas, training is immediately halted. Failures observed include silent data corruption from broken GPUs and gradient corruption due to race conditions.

The final model, Laguna S (118B parameters), was trained on 30 trillion tokens across 4,000 GPUs. Early results show strong performance in agentic coding benchmarks (e.g., SWEETbench), outperforming models like GLM 4.5 Air.

POCs are often temporary, unmeasured, and reversible experiments that fail because they lack a defined path to production ownership. This creates an 'AI zombie' state where the system exists but delivers no measurable value or transformation.

Successful deployment requires addressing technical, organizational, and cultural gaps. The biggest hurdles are unclear ownership (who owns it when it breaks?), lack of dedicated funding for operations, and a culture that rewards demos over deployment frequency.

To escape the POC prison, implement a structured 90-day program: (1) Document current reality (ugly processes/data); (2) Achieve commitment and accountability by selecting one initiative; (3) Deploy into real systems under supervision to build the paved road.

Compliance and governance cannot live in slide decks or meetings. They must be embedded directly into the delivery process (e.g., 'governance as code'), making rules executable, auditable, and non-negotiable.