Training 1T-Param Kimi K2.7 with RL in SkyRL: Two-Node (16x B300) 262K Full-Context Reference

Image source: https://github.com/moonshotai/kimi-k2

On October 3, 2026, developer Casper Hansen (@casper_hansen_) highlighted a reference setup demonstrating that Kimi K2.5–K2.7 models can be trained with reinforcement learning (RL) on just 16 B300 GPUs (two nodes) at full context. Developed in collaboration with SkyRL maintainers, the recipe challenges conventional post-training setups that typically require 16 full nodes without guaranteeing full-context support.

The significance is straightforward: post-training reinforcement learning for 1-trillion-parameter mixture-of-experts (MoE) architectures has traditionally demanded large clusters. This reference provides a concrete, documented setup for training at a full 262,144-token context within a realistic two-node footprint.

Inside the Kimi K2 INT4 LoRA Recipe

The underlying setup is detailed in the official SkyRL documentation example, Kimi K2.x: 1T-param RL with INT4 rollouts + LoRA (docs.skyrl.ai/docs/examples/kimi_k2_int4_lora). Key specifications include:

• Target Model: moonshotai/Kimi-K2.7-Code, a 1T-parameter MoE architecture based on DeepSeek-V3 (61 layers, 384 routed experts, MLA). Expert weights use INT4 QAT quantization (compressed-tensors pack-quantized, group_size=32).
• Hardware Profile: Two 8-GPU nodes (verified on 2 x 8 B300). One vLLM engine per node colocated with training, configured with tensor parallelism (TP=8).
• Context Length: Full 262,144 tokens.
• Training Strategy: Combines DAPO-style GRPO, LoRA, and bit-exact fake-INT4 quantization-aware training (QAT) to minimize memory consumption and communication overhead.
• Required Backend: Uses the megatron training backend.

Practical Audience, Workflows, and Constraints

SkyRL (skyrl-train) is a modular open-source RL framework designed to simplify prototyping training algorithms, environments, and execution plans (including custom model placement, training/generation colocation, and asynchronous RL). It supports training backends across FSDP, FSDP2, Megatron, and DeepSpeed, alongside inference backends including vLLM, SGLang, and custom OpenAI API-compatible endpoints, supporting algorithms like PPO and GRPO.

• Target Audience: Researchers and infrastructure engineers seeking to post-train 1T-scale MoE models without multi-cluster scale.
• Usage Path: Reviewing the Kimi K2 INT4 LoRA example in the official SkyRL documentation provides the baseline configuration and execution flags.
• Key Constraints: The reference explicitly relies on high-bandwidth B300 GPU hardware, the Megatron backend, INT4 QAT rollouts, and LoRA. Memory placement and TP=8 configuration across vLLM engines and Megatron must be configured precisely to avoid out-of-memory errors during long-context rollouts.

For teams operating without massive cluster capacity, this documentation establishes a concrete reference baseline for what two nodes can accomplish. While sixteen B300 GPUs remain substantial infrastructure, it shifts 1T MoE full-context RL from an exclusive massive-cluster task into a documented recipe for small multi-node teams.

Sources

• SkyRL Documentation: Kimi K2.x: 1T-param RL with INT4 rollouts + LoRA
• Casper Hansen (@casper_hansen_) on X (2026-10-03): Kimi K2.5-2.7 RL on 16 B300 GPUs