Coding-agent research
Study reinforcement learning for long-horizon repository work with published weights, data, logs and training code.
Independent tool overview
Deep SWE at a glance
DeepSWE-Preview is a 32-billion-parameter open-weight coding-agent model created by Agentica and Together AI on top of Qwen3-32B. It is a research release for repository-scale software-engineering tasks, not a hosted coding app or a current Together AI serverless model.
Visit the official Deep SWE site ↗
- Release
- DeepSWE-Preview, July 2025
- Base model
- Qwen3-32B with thinking mode
- Creators
- Agentica and Together AI
- License
- MIT on the Hugging Face model card
- Access
- Downloadable model weights; self-hosting or a compatible inference provider required
- Launch benchmark
- 42.2% Pass@1; 59% with hybrid best-of-16 selection on SWE-bench Verified
Overview
What Deep SWE is
DeepSWE-Preview was trained with reinforcement learning to inspect repositories, search files, edit code, run shell commands and submit a patch. The team released the model weights, a 4,500-task training subset, training code and evaluation logs, making it most useful to researchers and engineering teams building their own coding-agent stack.
At launch, the authors reported 42.2% Pass@1 on SWE-bench Verified averaged across 16 runs. The headline 59% result used hybrid test-time scaling to generate and select among 16 candidate trajectories, so it should not be read as the success rate of one ordinary run. The model card recommends temperature 1, 32K–64K output capacity and the R2E-Gym prompt and tool scaffold.
The weights remain downloadable under an MIT license, but DeepSWE-Preview is still labeled Preview and requires substantial infrastructure and agent plumbing. Teams that want an everyday coding assistant will generally be better served by a maintained hosted agent; DeepSWE is stronger as a reproducible research baseline or a starting checkpoint for further work.
Use cases
Who Deep SWE is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Custom self-hosted agents
Build a controlled coding system around an open 32B checkpoint and your own sandbox, tools and review gates.
Reproducible evaluation
Re-run or extend the published R2E-Gym and SWE-bench experiments instead of relying only on a closed product score.
Capabilities
Core Deep SWE features
Repository navigation
The training scaffold lets the agent search a codebase and inspect relevant files across a multi-step task.
File editing
DeepSWE can create and modify files through the R2E-Gym file-editor interface.
Shell execution
The agent can run commands and tests inside an isolated software-engineering environment.
Reinforcement-learned workflow
It was trained with outcome rewards based on whether generated patches passed selected tests.
Open research artifacts
The release includes weights, a filtered training dataset, training scripts, experiment logs and evaluation traces.
Multiple serving options
The model card documents Transformers, vLLM and SGLang paths for teams operating their own inference stack.
Process
How the Deep SWE workflow works
Step 1
Choose the use case
Treat DeepSWE as a model component. Define the agent scaffold, repositories, permitted commands and review boundary before deployment.
Step 2
Provision isolated inference and execution
Serve the 32B model on suitable GPU infrastructure and run generated commands in disposable, least-privilege sandboxes without production secrets.
Step 3
Match the documented scaffold
Start with the R2E-Gym system prompt and search, file-editor, shell and finish tools used by the published evaluation.
Step 4
Evaluate on representative tasks
Measure first-attempt pass rate, regressions, latency, token use and cost on your own repositories rather than importing the launch headline.
Step 5
Require human review
Inspect the diff, run the full test and security suite, and merge only through the same protections used for human contributions.
Cost
Deep SWE pricing and free plan
The DeepSWE-Preview weights and associated open-source artifacts do not carry a software subscription fee. Real cost comes from GPU inference, storage, sandboxes, repeated candidate runs and engineering operations. Together AI does not currently list DeepSWE-Preview in its public serverless model catalog, so there is no official per-token DeepSWE endpoint price.
Model weights
$0 license fee
Download the MIT-licensed checkpoint from Hugging Face.
- Infrastructure not included
- A capable inference server is required
- Review the licenses of the base model, dependencies and data for your use case
Self-hosted inference
Variable GPU cost
Run with Transformers, vLLM, SGLang or another compatible stack.
- A 32B model generally requires substantial GPU memory
- Quantization and parallelism change performance and quality
- Sandbox compute and storage are additional
Together AI infrastructure
From published GPU rates
Use dedicated inference or GPU clusters where supported; this is infrastructure pricing, not a DeepSWE serverless SKU.
- On-demand H100 dedicated inference is listed at $5.49 per GPU-hour
- On-demand H100 GPU clusters are listed at $3.99 per GPU-hour
- Actual deployment may require multiple GPUs and other services
Pricing checked . Check current pricing at the source ↗
Assessment
Deep SWE strengths and limitations
Where it stands out
- Publishes substantially more of the training and evaluation stack than a typical coding-agent release.
- Targets multi-step repository work rather than isolated code completion.
- The 32B checkpoint is smaller and more approachable than many frontier-scale open models.
- Launch results distinguish single-run performance from best-of-many selection in the underlying report.
- MIT-licensed model weights give teams room to inspect, adapt and self-host the release.
What to consider
- DeepSWE-Preview is a model checkpoint and research scaffold, not a polished editor extension or managed coding service.
- The 59% launch headline depends on 16 generated trajectories plus hybrid selection; one-run Pass@1 was 42.2% in the authors' evaluation.
- Launch-time state-of-the-art language is no longer a current market comparison and newer coding models may outperform it.
- The public Together serverless catalog does not currently list a DeepSWE-Preview endpoint.
- Serving a 32B reasoning model with long outputs can be expensive and operationally demanding.
- Benchmark environments do not capture every dependency, security constraint or failure mode in a private production repository.
- Generated shell commands and patches can damage data, introduce vulnerabilities or expose secrets without sandboxing and review.
Compare
Deep SWE alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Coding
Qwen3-Coder-Next
Qwen3-Coder-Next is a newer open-weight coding model designed for agentic workloads and may be a more current self-hosted baseline.
Explore Qwen3-Coder-Next →Coding
Devstral 2
Devstral 2 is a newer coding-focused model family for teams comparing open agentic coding checkpoints.
Explore Devstral 2 →Coding
Cursor Agents
Cursor Agents is a managed product for teams that want a usable remote coding workflow without building the model and sandbox stack themselves.
Explore Cursor Agents →Questions
Deep SWE FAQs
Is DeepSWE a coding app?
No. DeepSWE-Preview is an open-weight 32B model and research release. You need to supply inference infrastructure, an agent scaffold, repository access, tools and sandboxing.
Is DeepSWE free?
The model weights are available under an MIT license with no subscription fee. GPU inference, storage, execution sandboxes and operations still cost money.
Can I use DeepSWE through the Together AI API?
DeepSWE-Preview is not listed in Together AI's current public serverless catalog. You can self-host the weights or investigate compatible custom and dedicated deployment options.
What does the 59% SWE-bench score mean?
It is the authors' 2025 SWE-bench Verified result using hybrid test-time scaling to choose among 16 trajectories. Their average single-run Pass@1 result was 42.2%.
What settings does the model card recommend?
The model card recommends temperature 1, at least 32K–64K maximum output tokens and the R2E-Gym prompt and tools used during evaluation.
Is DeepSWE safe to run on a production repository?
Not without controls. Use an isolated environment, least-privilege credentials, command restrictions, full automated tests and mandatory human review before any merge or deployment.
Bottom line
Our Deep SWE verdict
DeepSWE-Preview is still a useful open research artifact for teams studying or building coding agents, especially because its weights, data, training code and logs are available together. It is not the simplest route to a production coding assistant, and its launch benchmark headline needs the best-of-16 context. Choose it when openness and experimentation matter more than turnkey usability.
Visit Deep SWE website ↗Similar Coding Tools
View all Coding AI tools →
Bolt.new
Bolt.new: Lets you spin up AI-powered web apps instantly using natural language.
Emergent 2.0
Emergent 2.0 - Agentic vibe coding platform for production-ready apps
Gemini CLI
Gemini CLI - Google's open-source terminal agent with high free usage limits
Devstral
Devstral - Mistral's open model family for agentic coding
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.