Platform Comparison

Previous
  • Multi OSS Training Backbone: Support switching between multiple open-source training backbones quickly.
  • Multi OSS Infer Backbone: Support both vLLM and SGLang.
  • Low Code Change: Do not require too many edits to convert a user‑defined (multi) agent workflow into trainable workflows.
  • Without-GPU (Cloud-Computing): Rollout and power RL training in a laptop without GPU, using Tinker (AgentLightning) or without Tinker (AgentJet-TinkerScript, comming soon)
  • Timeline Optimization: Automatically merge shared-history context generated by the same agents to promote training speed.
  • Open Bench Platform: Trace baseline environment's performance across git history in different training backbones.
  • Multi-Agent Optimization: Deal with sophisticated multi-agent interaction efficiently, automatically clustering and merging samples generated by the same agents.
  • High-res Rollout Logging: Integrated with token level rollout trajectory, highlighting token logprob and loss mask for deep‑level research.
  • Agentic Framework Compatible: Easy to convert AgentScope and Langchain workflows into trainable workflows.
Feature AgentJet (Beta) AgentLightning rLLM VeRL 0.7.0
Multi OSS Training Backbone ✅ ❌ ❌ ➖
Multi OSS Infer Backbone ✅ ✅ ✅ ✅
Low Code Change ✅ ✅ ✅ ➖
Without-GPU (Cloud-Computing) (Testing) ✅ ❌ ➖
Timeline Optimization ✅ ❌ ❌ ➖
Open Bench Platform ✅ ❌ ❌ ✅
Multiagent Optimization ✅ ❌ ❌ ✅
High-res Rollout Logging ✅ ❌ ❌ ❌
AgentScope Compatible ✅ ✅ ✅ ➖
Langchain Compatible ✅ ✅ ✅ ➖
  • ✅ = "supported"; ❌ = "not supported"; ➖ = "not applicable"
  • All projects are quickly evolving. We expect features not supported today will catch up sooner or later.
Tracing Feedback