Robots Atlas>ROBOTS ATLAS
Agents

DeepSWE

2025ActivePublished: 6 August 2026Updated: 6 August 2026Published
Key innovation
Demonstrating that a strong software-engineering agent can be trained purely with reinforcement learning (RL) on top of an open model (Qwen3-32B), without supervised fine-tuning (SFT), reaching SOTA results among open-weight agents.
Category
Agents
Abstraction level
System
Operation level
ModelApplicationTooling
Use cases
Autonomously solving software-engineering tasksAn open, reproducible base for RL research on coding agentsA reference point for open-weight agents on SWE-BenchSelf-hosting a coding agent (open weights)Fine-tuning / building custom RL agents

How it works

The starting point is the open Qwen3-32B model. Instead of SFT, reinforcement learning is used (the rLLM framework): the agent solves coding tasks in an execution environment, and the reward signal comes from verifiable correctness (e.g. passing tests / solving a SWE-Bench task). The model learns to plan, edit code and verify results over many steps. Test-time scaling is additionally used, e.g. generating and selecting among multiple solutions, which raises success (Pass@16, 59% with test-time scaling).

Problem solved

Strong coding agents usually require costly supervised fine-tuning on large demonstration datasets. DeepSWE shows they can be obtained purely via RL from an open model, fully openly and reproducibly.

Components

Base model (Qwen3-32B)The starting point (backbone).

The open language model from which agent training starts.

Official

RL training (rLLM, no SFT)The agent's learning mechanism.

Pure reinforcement learning in the rLLM framework, without supervised fine-tuning.

Verifiable reward (tests/SWE-Bench)The RL optimization target.

A reward signal from objective verification of solutions (passing tests).

Test-time scalingImproving results without changing weights.

Generating and selecting among multiple solutions at inference for higher success.

Official

Implementation

Implementation pitfalls
Dependence on verifiable rewardsMedium

RL needs a reliable, objective signal (e.g. tests); for tasks without clear verification the approach is harder.

Fix:Designing environments with automatic, reliable verification and avoiding reward hacking.
RL compute costMedium

Training an RL agent in execution environments is computationally and engineering-wise expensive.

Fix:Efficient frameworks (rLLM), distributed environments and careful selection of training tasks.

Evolution

2025
DeepSWE-Preview release (Agentica x Together AI)
Inflection point

On July 2, 2025 an open coding agent trained purely with RL from Qwen3-32B was released, with a full training recipe and weights; SOTA among open-weight agents on SWE-Bench-Verified.