AI Models
SWE-bench Verified is a 500-instance, human-validated subset of SWE-bench used to evaluate coding agents on real GitHub issue-resolution tasks. Open-weight, permissively licensed models are now achieving performance that rivals proprietary systems, but deploying these models at scale requires robust AI infrastructure.
SWE-bench Verified is a 500-instance, human-validated subset of SWE-bench used to evaluate coding agents on real GitHub issue-resolution tasks. Open-weight, permissively licensed models are now achieving performance that rivals proprietary systems, but deploying these models at scale requires robust AI infrastructure. This guide examines six leading open-weight coding models alongside Modal, which provides serverless AI infrastructure for inference, training, batch jobs, and secure Sandboxes for executing AI-generated code. Whether you need parameter-efficient models for local deployment or large-scale reasoning models for complex software engineering tasks, this breakdown covers the essential options for 2026.
Modal provides the serverless compute foundation for deploying and evaluating the open-weight coding models that follow. While the other entries in this guide are AI models, Modal is the AI infrastructure platform where teams serve coding agents in production, run model-evaluation workflows, and scale GPU-accelerated inference without managing clusters.
Modal has completed SOC 2 Type II and supports HIPAA-compliant workloads on Enterprise plans via a BAA. The platform uses TLS 1.3 for public APIs and encryption for data in transit and at rest.
Modal powers cloud infrastructure for over 10,000 teams, including AI companies building coding agents at scale. The platform's code-first SDK eliminates YAML configuration, with code-defined infrastructure available in Python, TypeScript, and Go, enabling teams to define compute requirements directly in code.
Best For: Teams deploying open-weight coding models to production, running model evaluations at scale, or building coding agents that need secure execution environments with on-demand GPU access.
Qwen3-Coder-Next represents the efficiency breakthrough of 2026, achieving top-tier SWE-bench performance while activating a fraction of its total parameters. This Mixture-of-Experts model from Alibaba's Qwen team makes local deployment viable on high-memory machines.
Qwen3-Coder-Next achieves 70.6% on SWE-bench Verified and 44.3% on SWE-bench Pro per its official model card.
Qwen3-Coder-Next is designed for local development and has 80B total parameters with 3B activated per token. Quantized deployments may run on high-memory consumer or workstation-class machines, but exact memory requirements depend on quantization, runtime, and KV-cache settings. This makes it one of the more accessible high-performance coding models for developers without cloud GPU access.
Best For: Teams seeking high SWE-bench performance achievable on local or workstation hardware, or cloud deployments where inference cost matters due to low active parameter count.
DeepSeek-V3.2 brings the largest parameter count in this guide, combining advanced reasoning capabilities with strong coding performance. DeepSeek reports that DeepSeek-V3.2-Speciale, its high-compute reasoning variant, achieved gold-medal-level performance in the 2025 International Mathematical Olympiad and International Olympiad in Informatics.
DeepSeek-V3.2 reports 70% on SWE-bench Verified on its official model card. Higher figures such as 73.1% require stating the exact evaluation setup and citing the source that reports it. The model's 15.56% on SWE-bench Pro with the public dataset shows room for improvement on harder tasks.
DeepSeek-V3 lineage models have been documented with 128K context windows, smaller than Qwen3-Coder-Next's 256K, which may matter for extremely large codebases. However, the model's reasoning capabilities make it effective at understanding complex code relationships within its context limit.
Best For: Teams tackling complex software engineering problems that require strong reasoning, particularly where the coding task involves algorithmic challenges or mathematical reasoning.
Devstral-Small-2-24B-Instruct-2512 from Mistral AI adds vision support to Mistral's agentic coding model lineup, enabling agents to understand screenshots, diagrams, and visual documentation alongside code.
The model reports 68.0% on SWE-bench Verified and 55.7% on SWE-bench Multilingual. For reference, the larger Devstral 2 (123B) variant reaches 72.2% on SWE-bench Verified.
Mistral AI describes Devstral Small 2 as excelling at "using tools to explore codebases, editing multiple files and powering software engineering agents." The vision capability opens use cases unavailable to text-only models.
Best For: Teams building coding agents that need to process visual information alongside code, such as UI development assistants or agents that reference documentation with diagrams.
DeepSWE-Preview, from Agentica / Berkeley Sky Computing Lab / BAIR with Together AI collaboration, takes a fundamentally different approach: training entirely with reinforcement learning without any supervised fine-tuning step. This pure-RL methodology demonstrates that high-quality coding agents can emerge from reward-driven learning alone.
DeepSWE-Preview achieves 42.2% Pass@1 on SWE-bench Verified averaged over 16 runs. With hybrid test-time scaling, performance reached 59%; at its July 2025 release, the authors described this as leading open-source SWE-bench performance.
The model was trained on 64 H100s over 6 days using 4.5K problems from the R2E-Gym subset, filtered to avoid SWE-bench contamination. The base model is Qwen3-32B with thinking mode enabled.
Best For: Research teams exploring RL-based training approaches, or production deployments that can leverage test-time scaling for higher accuracy on critical tasks.
SWE-Swiss-32B takes a structured approach to software engineering tasks, decomposing issue resolution into three core skills: Localization, Repair, and Unit Test Generation. This multi-task methodology reports strong performance comparable to models twice its size.
SWE-Swiss-32B reports 60.2% on SWE-bench Verified. The project documentation describes this "performance [is] comparable to models more than double its size." Additional primary sources include the GitHub repository and the ACL Anthology entry.
The task decomposition approach provides interpretability advantages. Teams can analyze where their coding agent struggles, whether localization, repair, or test generation, and target improvements accordingly.
Best For: Teams wanting a structured, interpretable approach to coding agents, particularly those building systems that need to explain their reasoning or target specific improvement areas.
SWE-agent-LM-32B is created by the SWE-bench team (Princeton/Stanford NLP), representing the team's approach to training coding agent models using high-quality synthetic data from SWE-smith.
The model achieves 40.2% on SWE-bench Verified using the SWE-agent scaffold. At release, it represented the open-source state-of-the-art, though subsequent models have surpassed this score.
SWE-smith provides 50K task instances, an order of magnitude larger than prior open-source training sets for coding agents. This scale demonstrates the value of synthetic data generation for specialized tasks.
Best For: Teams wanting a baseline model from the benchmark creators, or those building on the SWE-agent scaffold and seeking a model optimized for that specific framework.
Choosing the right model is only half the challenge. Deploying these models at scale requires infrastructure that handles the unique demands of coding agents. Modal's platform addresses the core infrastructure requirements:
Coding agents generate and execute code autonomously, making sandboxed execution critical. Modal Sandboxes provide gVisor-isolated containers that Modal describes as supporting 100k+ concurrent sandboxes, with observability for monitoring agent behavior.
Large models like DeepSeek-V3.2 require significant GPU resources. Modal's broad GPU lineup lets teams match compute to model requirements, from L4s for smaller models to H100s and B200s for the largest deployments.
Teams customizing these open-weight models can use Modal Training, which supports single-node experiments and multi-node GPU training with B200/H200/H100 clusters.
Modal Sandboxes support secure execution of untrusted or agent-generated code and can be used for test-suite-style evaluation workflows. Modal describes Sandboxes as supporting 100k+ concurrent sandboxes, providing isolated environments for running test suites at scale.
Modal describes its Core Platform as flexible infrastructure for AI workloads, with an AI-native container runtime, storage, networking, and data primitives, observability, and multi-cloud capacity. This translates to faster cold starts, more efficient GPU utilization, and lower operational overhead compared to general-purpose cloud infrastructure.
Modal powers cloud infrastructure for over 10,000 teams, with coding agent deployments including Ramp, which uses Modal to power Ramp Inspect, a background coding agent that writes over half of all merged pull requests at Ramp. This production track record demonstrates enterprise-scale reliability.
SOC 2 Type II certification and HIPAA support address enterprise compliance requirements. Modal supports HIPAA-compliant workloads on Enterprise plans via a BAA. Modal Sandboxes are built on gVisor and are designed to limit the blast radius of malicious code to the Sandbox container, protecting both the infrastructure and other workloads.
Modal's code-first SDK enables code-defined infrastructure with no YAML, with support for Python, TypeScript, and Go. Teams define compute requirements directly in code, enabling rapid iteration as coding agent architectures evolve.
For teams deploying the open-weight models in this guide, whether Qwen3-Coder-Next for parameter-efficient inference or DeepSeek-V3.2 for reasoning-heavy tasks, Modal provides the infrastructure foundation that makes production deployment practical.
Explore the Modal documentation to get started with coding agent deployment.
Explore the Modal documentation to get started with coding agent deployment.
View Modal DocsSWE-Bench-style coding agents are AI systems designed to solve real software engineering tasks, specifically GitHub issues from open-source projects. The benchmark evaluates whether an agent can understand an issue description, locate relevant code, generate a fix, and produce a patch that passes the project's test suite. These agents combine language model capabilities with tool use, code search, and execution abilities.
Open-weight, permissively licensed models offer transparency, customization potential, and freedom from vendor lock-in. With licenses like Apache 2.0 and MIT, teams can fine-tune models on proprietary codebases, deploy them in air-gapped environments, and audit model behavior. Models like Qwen3-Coder-Next achieving 70.6% on SWE-bench Verified demonstrate that open-weight alternatives now compete with proprietary systems.
Qwen3-Coder-Next prioritizes parameter efficiency, activating only 3B of its 80B total parameters per token, enabling local or workstation-class deployment in quantized configurations. DeepSeek-V3.2 emphasizes reasoning capabilities with 685B parameters; DeepSeek reports that the high-compute DeepSeek-V3.2-Speciale variant achieved gold-medal-level performance in IMO and IOI settings. Choose Qwen3-Coder-Next for cost-efficient inference; choose DeepSeek-V3.2 for complex reasoning tasks.
Coding agents execute AI-generated code autonomously, requiring sandboxed execution to prevent malicious or buggy code from affecting systems. Key considerations include container isolation (gVisor, Firecracker), network segmentation, resource limits, and audit logging. Modal's security practices address these through SOC 2 Type II compliance, HIPAA support, and gVisor-based sandboxing. Modal supports HIPAA-compliant workloads on Enterprise plans via a BAA.
Current models excel at specific tasks such as fixing bugs, implementing features from clear specifications, and writing tests, but full SDLC automation remains aspirational. Models in this guide report roughly 40% to low-70% SWE-bench Verified performance depending on model, scaffold, and evaluation setup, meaning they solve real issues but not consistently. Human oversight remains essential for architecture decisions, security review, and complex requirements analysis.
Modal provides serverless infrastructure specifically optimized for AI workloads. Teams deploy coding models through Modal Inference with fast cold starts and automatic scaling. Secure code execution runs in Modal Sandboxes, which Modal describes as supporting 100k+ concurrent sandboxes. For fine-tuning, Modal Training offers multi-node GPU clusters. Modal provides infrastructure primitives, including Sandboxes, GPUs, batch jobs, and inference endpoints, that can support model-evaluation workflows.