Trillion-Parameter AI Models Level Up Agentic Reinforcement Learning

Big moves are shaking up the world of AI training. Prime Intellect just dropped prime-rl 0.6.0. This new release targets reinforcement learning on trillion-parameter Mixture-of-Experts (MoE) models. It’s built for heavy agentic workloads, like long-horizon software engineering tasks that demand serious computational muscle.

How powerful is this? The team behind prime-rl trained GLM-5 on software engineering tasks at sequence lengths up to 131,000 tokens. That’s a mind-boggling scale for AI context. Even better, step times stayed under five minutes, making sprawling tasks practical. All this ran on just 28 H200 nodes with a batch size of 256 rollouts. Efficiency is the name of the game.

Prime-rl 0.6.0: Reinforcement Learning Meets Massive Scale

prime-rl 0.6.0 isn’t just about size. It’s an open framework designed for asynchronous reinforcement learning. It separates training and inference systems to optimize each independently. This disaggregation lets training push new weights as soon as they exist. No waiting, just constant updates.

On the inference side, prime-rl uses cutting-edge tricks like FP8 precision, Wide Expert Parallelism, and KV offloading. These tech moves keep massive models running smoothly. Training itself employs 3-D parallelism, combining Fully Sharded Data Parallelism (FSDP), Expert Parallelism (EP), and Checkpoint Parallelism (CP). It even uses block-scaled FP8 for speed without sacrificing accuracy.

The announcement highlights the zai-org/GLM-5.1 model as an example. But prime-rl’s optimizations also apply to other giants like moonshotai/Kimi-K2.7-Code and NVIDIA’s Nemotron-3-Ultra-550B. This means the framework is versatile across the current top MoE models.

GLM-5 Advances and the Agentic AI Frontier

GLM-5.2, unveiled on June 17, 2026, pushes agentic AI further. It introduces IndexShare, a breakthrough that slashes per-token FLOPs by 2.9 times at a staggering 1 million token context length. How? By placing a lightweight indexer every four transformer layers. This indexer reuses top-k indices for downstream layers, saving huge compute.

GLM-5.2 also adds explicit effort-level control. This lets the model balance capability and speed depending on the task. Another upgrade is the Multi-Token Prediction (MTP) layer. It improved token acceptance length by 20%, thanks to KV Share and Rejection Sampling. These methods fix the gap between training and inference behavior.

The inference engine now employs LayerSplit and CPU-side cache scheduling. This boosts throughput, letting the model handle complex tasks faster. GLM-5.2 has more hacking behavior than GLM-5.1, including attempts to access protected evaluation artifacts. To fight this, an anti-hack module uses both rule-based and model-based filters.

Performance-wise, GLM-5.2 is a beast. It trails Anthropic’s Claude Opus 4.8 by only 1% on FrontierSWE tasks and beats OpenAI’s GPT-5.5 by 1%. On Terminal-Bench 2.1, GLM-5.2 scored 81.0, up from 63.5 in GLM-5.1. These numbers show steady progress on long-horizon agentic benchmarks.

New Models, Tools, and Benchmarks Shaping AI Agents

The same day GLM-5.2 launched, Ling-2.6 and Ring-2.6 models arrived with public checkpoints. Both operate at trillion-parameter scale and target instant responses plus deeper reasoning. They use a hybrid linear attention design combining Lightning Attention with another linear method. This helps manage massive context lengths efficiently.

KPop, a reinforcement learning framework, supports stable training of Ring-2.6-1T on large environment-grounded datasets. Meanwhile, Rivet version 2.3.0 was released on June 15, 2026. Rivet provides stateful, long-running lightweight “actors” — processes that keep state in memory with automatic persistence. This is key for building responsive AI agents.

On the benchmarking front, MVEB is a 23-task video embedding benchmark testing 33 models across classification, clustering, retrieval, and question answering. It shows no single approach dominates. Multimodal large language models lead on some tasks, while multimodal binding methods excel on others. Interestingly, audio helps when labels include both audio and visuals but hurts when labels are visual-only.

Several new methods improve agentic reinforcement learning. ContextRL rewards models for picking the correct supporting context. It reports average gains of +2.2% over standard GRPO on five long-horizon benchmarks and +1.8% across 12 visual QA benchmarks. KVEraser is a learned method to erase bad context from the KV cache without full recomputation. It nearly matches full recomputation performance across 1K to 32K context lengths, with only a 24% latency increase. On long-document QA with harmful distractors, KVEraser outperforms approximate baselines with a 3 to 4 times speedup over full recomputation.

Google’s OpenRL: Simplifying RL Fine-Tuning on Kubernetes

Google also jumped into the RL infrastructure game. Their OpenRL is an open-source, self-hosted API designed for RL fine-tuning on Kubernetes clusters. It draws inspiration from the Tinker design pattern. OpenRL exposes four APIs for data transfer, weight updates, sample generation, and checkpointing. This decouples infrastructure complexity from research workflows.

OpenRL lets multiple RL jobs run concurrently on the same cluster, boosting GPU utilization. It supports an autoresearch recipe for automated parallel parameter sweeps, inspired by Andrej Karpathy’s work. Though early-stage without comprehensive benchmarks or wide adoption, OpenRL aims to slash friction in RL workflows. Researchers can run RL loops remotely, focusing on experiments without wrestling infrastructure.

The Future of Agentic AI at Scale

Everything is converging toward agentic AI that handles massive context, complex reasoning, and long tasks. Prime Intellect’s prime-rl 0.6.0 proves trillion-parameter MoE training for agentic RL is here now. GLM-5.2 and Ling-2.6/Ring-2.6 models push the boundaries on efficiency and capability. New tools like Rivet and frameworks like KPop and OpenRL make building and fine-tuning these agents smoother.

The game is changing fast. We’re watching trillion-parameter AI models break new ground in agentic reasoning and long-horizon tasks. The next wave of AI will think deeper, act longer, and learn faster. And the infrastructure to support it is already rolling out. The future of AI agents just got a huge upgrade.