Our run built successfully, but tests failed
Our run on October 13, 2026, at commit a025e03 recorded successful install and build steps, but the test command failed. The full recorded commit is a025e0373addc2cd76e77a5064d60c517f05dcf8. These observations come from our stored sandbox record; this correction involved no new execution.
The install, build and test steps took 106, 196 and 72 seconds respectively. Our harness records each command's success flag, exit code, elapsed time and timeout status, alongside parsed output. Here, install and build returned zero, while tests returned one without timing out. A zero build exit does not establish that a model loaded, an inference request completed or a deployment worked.
The attached stored parser summaries report 48 passing and 3 failing package results, with a total of 51. They do not establish exhaustive unique test coverage. The supplied tail repeats package results under a summary section, so counting those lines again would inflate the evidence. No complete-log reconciliation was performed, and the supplied record does not establish a skipped-test count.
Separately, the raw tail names TestStart_GpuRestartWaitResetsAfterStats in internal/perf, with the message timed out waiting for GPU stats from collector 6. That identifies one visible failure; it does not explain every failure in the stored parser summary or establish the underlying cause. Entries marked [no test files] are not failed tests.
We did not test inference quality, tokens per second, model-loading delays, memory savings, GPU drivers or an exposed service. The record contains no audit result or scanned-package coverage, so it supports no security conclusion. The testing-methodology page explains the publisher's recording approach and limits; its current text does not prove that this older run used today's procedure.
What the proxy takes over
llama-swap selects an upstream server from the requested model ID and replaces the wrong running upstream. In the basic configuration, it handles one model at a time; the pinned README documents a matrix for concurrent models. It also documents idle unloading through ttl, startup hooks, aliases and manual unload controls.
That makes it useful when your client should keep one endpoint while different configured servers handle requests. The pinned README lists OpenAI and Anthropic routes, along with upstream-specific routes for several text and media backends. Those documented routes are capabilities to validate with your client and backend, not compatibility results from our sandbox.
The operator still supplies the upstream command and model path. llama-swap's lifecycle controls do not establish that your chosen quantization, context size or GPU configuration fits the machine. Start by checking one backend directly, then route the same request through the proxy before adding more models.
Scheduling changed after the measured commit
Issue 1004 was closed as completed on October 24, 2026, after capacity-based routing work. The issue discussion describes resource pools, configured model capacities and eviction preferences. It therefore no longer supports describing capacity scheduling as merely an outstanding request.
That work postdates our October 13 run at a025e03. Its completion commit, 957cc76, is not our measured commit. The supplied evidence does not establish that our run exercised the completed capacity router, and it does not establish automatic measurement of model resource requirements. Check the configuration documentation for the version you intend to deploy.
Check streaming through your actual endpoint
The pinned README recommends disabling Nginx response buffering for streaming endpoints. It also recommends explicit buffering configuration even though llama-swap sends X-Accel-Buffering: no on SSE responses. Test streaming through the address your client will use, including any reverse proxy.
API-key support, logs and request inspection are documented, but our run did not validate a public deployment or access-control policy. Decide who can reach the API and interface, and check what request data your deployment exposes through logs.
When to choose a different service
Ollama provides a REST API for running and managing models. Consider Ollama when that model-running and management workflow answers your need without configuring a proxy around separate upstream commands. We did not compare its installation effort or runtime performance with llama-swap.
LocalAI pulls separate backend images when a model needs them. Consider LocalAI when you want the service to manage those backend components. Its supplied README describes a modular system, so calling it a bundled collection of every backend would misrepresent the current design. These are alternative operating approaches, not measured winners.
How to read the editorial scores
Scores describe the narrow dimensions named here, not an overall production-readiness rating. The scale is 1 for substantial evidenced obstacles, 2 for major operator work, 3 for a workable path with material gaps, 4 for strong evidence with a specific limitation, and 5 for comprehensive evidence within the stated dimension. Community assesses the substance of the supplied public collaboration, not adoption or support response times. Maturity assesses the documented operational controls and recorded development checks, not deployment reliability.
The setup score reflects documented binary and configuration options plus successful install and build steps, with upstream setup still required. Documentation covers installation, configuration, routes, shutdown and streaming caveats. The community evidence includes a detailed scheduling discussion involving users, contributors and the owner, followed by completed work. Maturity receives credit for lifecycle and monitoring controls, but the failed test command remains an unresolved limitation.

