NVIDIA 发布 SWE-Serve：评测编码智能体在 SGLang 推理服务任务上的表现

Technical Blog

sessionStorage.removeItem('facets.search.text'); sessionStorage.removeItem('facets.search.text-copy'); sessionStorage.removeItem('facets.search.filters');

Sep 23, 2026

SWE-Serve evaluates AI coding agents on 53 inference-engineering tasks derived from 83 merged SGLang pull requests across six families including model enablement, decoding, caching, scheduling, serving APIs, and distributed execution. Nineteen tasks start a live server and test patches through the full serving path; across these tasks the same patches pass 45.9% of the time with complete verification but 69.4% when live-serving checks are excluded, meaning about one in three patches that pass other tests fail live-serving validation. Tasks spanning multiple runtime domains such as request handling, scheduling, model execution, and KV-cache management show a 21.3 percentage-point lower pass rate than single-domain tasks, with every tested model exhibiting the same gap. Eleven models evaluated with mini-swe-agent under closed-book conditions show mean pass@1 ranging from 34.6% to 75.5%; top performers include Claude Opus 5 and GPT-5.6 Sol at 75%, though cost and wall time vary widely among models with similar scores. Each task was validated on declared hardware with the unmodified repository failing new-behavior tests while passing regressions, and a reference patch passing the complete verifier; web access was blocked during evaluation to prevent retrieval of upstream solutions. Next Steps Explore the SWE-Serve leaderboard to compare model performance across inference-engineering tasks. Run SWE-Serve on GitHub to evaluate your coding agent on SGLang inference-engineering tasks. Read the SWE-Serve paper for the full qualification and evaluation-integrity process.

Powered by NVIDIA Nemotron. AI-generated content may summarize information incompletely. Verify important information. Learn more

An AI coding agent’s patch can pass tests yet fail when the server loads a real model and handles requests. Evaluating changes to inference-serving software therefore requires checking the full serving path, including whether the system returns correct results through its public interface.

Developed with input from the SGLang team, SWE-Serve evaluates this gap with 53 tasks derived from merged changes to SGLang, an open-source system for serving large language models. Across 19 tasks with live-serving checks, the same patches passed 69.4% of the time when those checks were excluded, but only 45.9% with the complete verifier. About one in three patches that passed the other checks failed live-serving tests.

Quick links: Read the paper | Explore the leaderboard | Run SWE-Serve on GitHub

Existing repository-level benchmarks evaluate coding agents across general software-engineering tasks, while inference benchmarks often concentrate on kernel generation or performance optimization. SWE-Serve instead tests repository-scale changes across the inference-serving stack, including model enablement, decoding, caching, scheduling, serving APIs, and runtime performance.

To evaluate this broader engineering work, SWE-Serve turns 83 merged SGLang pull requests into 53 executable tasks across six inference-engineering families.

Twelve tasks run on CPU, while 41 use a single NVIDIA H100. This first release doesn’t evaluate other inference engines, multi-GPU execution, or multi-node serving.

Thirty-seven tasks come from a single upstream pull request. The other 16 combine two to six related changes. In total, the benchmark draws on 83 merged SGLang pull requests. The SGLang team, a launch partner for SWE-Serve, contributed ideas for identifying challenging tasks, suggested particularly demanding pull requests, and helped shape our approach to verifying correctness.

Each task gives the agent an instruction and a containerized SGLang checkout from before the target change. The agent’s patch passes if it satisfies the task’s hidden verifier on the declared hardware. It is never compared with the reference implementation.

These are substantial changes. The median reference solution modifies 553 lines across seven files. A typical verifier has seven tests for the new behavior and 10 regression tests. Nineteen tasks start a real server, and three enforce a calibrated performance gate on an H100.

One task asks the agent to add serving support for dense and mixture-of-experts (MoE) Qwen3.5 models. Starting from a revision without Qwen3.5 support, the agent must make both the 0.8B dense model and the 35B-A3B MoE model load and serve through the normal SGLang interfaces on one H100.

Its verifier checks model registration, configuration and weight loading, image and video inputs, OpenAI-compatible requests, native batched generation, log probabilities, and execution through the MoE model’s routed experts.

Some failures only appear once a real server starts. The SGLang team contributed ideas for end-to-end verification, including recommendations for specific model-serving tests. SWE-Serve includes 19 tasks that load the required model and test the agent’s patch through a live serving interface.

Across these tasks, the same 627 patches pass 45.9% of the time under the complete verifier. When the live serving tests are excluded, that rate rises to 69.4%. In other words, 147 patches changed from fail to pass when the live serving tests were excluded.

The 19 tasks contain 276 live serving tests. Of those, 242 are sourced or adapted from SGLang. The remaining 34 cover behavior introduced by the corresponding merged changes when no suitable upstream test was available.

The Gemma 4 MoE task makes the result concrete. Sixteen of 33 patches passed every other check but failed at least one live serving test. Those tests cover model loading, expert routing, text and image serving, and batched generation with correct ordering and log probabilities.

Figure 1. Pass rates with and without live-serving tests (A) and by runtime-domain breadth (B)

A SWE-Serve pass has a narrow meaning: the patch satisfies the benchmark verifier. SWE-Serve’s tests are not SGLang’s upstream review process. They don’t establish that an agent patch or benchmark reference solution is deployable, ready to merge, or endorsed by SGLang maintainers.

We divided the request-to-output path into four runtime domains: request handling and I/O, scheduling and request lifecycle, model execution, and KV-cache and runtime-resource management.

Across the best setting for each of the 11 models, the 26 tasks confined to one runtime domain have a 69.0% pass rate. The 27 tasks spanning more than one runtime domain have a 47.7% pass rate, a difference of 21.3 percentage points. Every model setting shows the same direction of difference.

Model performance varies substantially. Across each model’s best tested configuration, mean pass@1 ranges from 34.6% to 75.5%. No model scores highest in every task category, and configurations with similar overall scores can have very different costs and runtimes.

We evaluated 11 models and 31 model-effort configurations with mini-swe-agent , a minimal software-engineering agent that uses only Bash, under closed-book conditions. We tested the two Claude and three GPT-5.6 models at five effort levels; the other six models ran at one setting each.

Each configuration ran the complete 53-task benchmark three times. Each agent session was capped at 210 minutes and 350 steps. A task counts as solved only when the agent’s patch passes the complete verifier on the task’s declared hardware. The table reports the highest-scoring effort setting for each model.

Native harnesses didn’t improve the two leaders: GPT-5.6 Sol scored 73.6% in Codex and Claude Opus 5 scored 69.8% in Claude Code, versus 75.5% each with mini-swe-agent .

Cost doesn’t map cleanly to performance. Among the four models tied at 64%, mean cost ranges from $0.95 to $7.24 per task, while mean wall time ranges from 25.5 to 99.9 minutes.

No model leads all six engineering families, and models with the same overall score can have different strengths. The top score shows that many SWE-Serve tasks are within reach of current agents; the spread shows that performance is far from uniform.

We screened 786 potential task sources, built 156 executable candidates, and admitted 53.

Every admitted task was tested on its declared hardware. The unmodified repository had to fail the tests for the new behavior while continuing to pass regression tests. A reference patch had to pass the complete verifier. We also challenged the verifiers with agent-created patches, repairing, narrowing, or excluding tasks when we found a concrete problem.

Reported evaluations are closed-book. We block the public web and upstream source repositories, while allowing Hugging Face access for model weights, because an open-network pilot showed models retrieving task-specific upstream code. We audited all 1,749 trials behind the leaderboard; 196 prohibited retrieval attempts were blocked, and none succeeded. The paper describes the full qualification and evaluation-integrity process.

SWE-Serve includes the task environments, verifiers, and baseline configurations. It makes the gap between passing local checks and working through the full serving path measurable.

Explore the SWE-Serve leaderboard , then run SWE-Serve on GitHub to evaluate your coding agent on SGLang inference-engineering tasks.