AgentIndex icon
AgentIndex
AReaL logo
AReaL
★ 5.9k
vs
dsh-anchored-standard logo
dsh-anchored-standard
★ 3.7k

AReaL vs dsh-anchored-standard

AReaL: AReaL is a fully asynchronous reinforcement learning system from Tsinghua IIIS and Ant Group for training large reasoning and agentic models. Version 2.0 splits it into training, inference, agent and weight-update microservices, and black-box agent apps can be trained online by pointing their base_url at AReaL.; dsh-anchored-standard: dsh-anchored-standard is a set of community DeepSeek Harness agent presets that start a session under the Minimal condition, then promote to a small resident tool catalog and unlock heavier Standard tools on demand. It was tuned for DeepSeek V4 Pro and was frozen on 2026-09-10 after that model's retirement; new issues and PRs are no longer accepted.

01

TL;DR

AReaL logo
AReaLLLM Infra · ★ 5.9k

RL fine-tuning of reasoning models on math and coding tasks

Read AReaL full review →
dsh-anchored-standard logo
dsh-anchored-standardDev Tooling · ★ 3.7k

Studying how API-visible tool surface affects agent trajectories

Read dsh-anchored-standard full review →
02

Side-by-Side Comparison

Field
AReaL logoAReaL
dsh-anchored-standard logodsh-anchored-standard
Category
LLM Infra
Dev Tooling
Stars
★ 5.9k
★ 3.7k
License
Apache-2.0
NOASSERTION
Install Command
git clone https://github.com/areal-project/AReaL && cd AReaL && pip install uv && uv sync --extra cuda
—
Updated
1d ago
2d ago
Open Source
Yes
Yes
Website
↗ Visit
↗ Visit
GitHub
↗ GitHub
↗ GitHub
Tags
agent, agent-framework, llm
agent-framework, deepseek, deepseek-harness
03

Features

AReaL logoAReaL
01Fully asynchronous RL training, reported 2.77x faster than synchronous systems in v0.3
02Online RL for black-box agents by replacing base_url and api_key
03Microservice architecture in 2.0 with separate training, inference, agent and weight-update services
04SGLang or vLLM inference backends, Ray clusters and Ascend NPU support
05Examples for math, SWE, search agents, tau2-bench and OpenClaw agents
dsh-anchored-standard logodsh-anchored-standard
01Base mode, two live-anchor variants and one seeded prefab mode
02Two-phase bootstrap: Minimal-aligned start, then full Standard tools
03context-gate unified injection control
04Published dose-response data on discovery tools
04

Use Cases

AReaL logoAReaL
↳RL fine-tuning of reasoning models on math and coding tasks
↳Training multi-turn agents such as search or customer-service agents
↳Continuously improving an existing agent app via online RL
dsh-anchored-standard logodsh-anchored-standard
↳Studying how API-visible tool surface affects agent trajectories
↳Reusing the context-gate and prefab pipeline in other harness setups
05

Best For

AReaL logoAReaL
EssentialTrending
dsh-anchored-standard logodsh-anchored-standard
—
FAQ

FAQ

What is the difference between AReaL and dsh-anchored-standard?
Both AReaL and dsh-anchored-standard are in the LLM Infra category. AReaL has 5.9k stars, while dsh-anchored-standard has 3.7k stars.
Which is better, AReaL or dsh-anchored-standard?
The best choice depends on your use case. Choose AReaL if RL fine-tuning of reasoning models on math and coding tasks, and dsh-anchored-standard if Studying how API-visible tool surface affects agent trajectories.
Is AReaL free or open source?
Yes, AReaL is open source on GitHub (Apache-2.0).
Is dsh-anchored-standard free or open source?
Yes, dsh-anchored-standard is open source on GitHub (NOASSERTION).
Cloud Infrastructure
Ad

Deploy on DigitalOcean

Host your MCP servers, Docker containers, and AI agents with $200 free credit over 60 days.

Claim $200 Free Credit
→

Related

Alternatives to AReaL →Alternatives to dsh-anchored-standard →AReaL details →dsh-anchored-standard details →