GLiClass vs Microsoft-Decision-1

GLiClass

An Apache-2.0 classifier that scores a whole label set in one encoder pass on the owner's hardware, with single-label, multi-label, hierarchical and few-shot modes. It returns label scores with no calibration claim, the bundled server has no authentication, and the last three test runs on the main branch, on 24 September 2026, failed.

Microsoft-Decision-1

Microsoft publishes $0.042 per million input tokens, with output free, and the model is callable through Microsoft Foundry and OpenRouter. It is in public preview. No licence, model-specific retention statement, rate limit or deprecation policy was found, and the Foundry route needs an Azure deployment and an Entra token. Latency and accuracy figures are Microsoft's claims.

GLiClass

1. Pass --host 127.0.0.1 to python -m gliclass.serve, or put the port behind your own gateway. The server checks no credential
2. On a machine without a GPU add --device cpu --dtype float32 --num-gpus-per-replica 0. The default configuration expects CUDA
3. Send one text a request to POST /gliclass. An array in texts is cut to its first item without an error
4. Set multi_label to false for one label from a set. The default scores each label independently, so scores do not sum to 1
5. Keep text plus labels under the pipeline's 1,024-token max_length, or use ZeroShotClassificationWithChunkingPipeline. Longer input is truncated silently

Microsoft-Decision-1

1. Use OpenRouter's Decisions method, not an OpenAI chat-completions SDK. OpenRouter says chat completions SDKs will not work with this model
2. On Foundry, send the request to the deployment's /providers/microsoft/v1/systemone path with a Microsoft Entra token for https://cognitiveservices.azure.com/.default, not an API key
3. Take the deployment name from the Foundry quickstart before the first call. Microsoft says to confirm the route and authentication header, and the pages reviewed do not give the name
4. Keep each request within OpenRouter's 32,768-token context and send only fixed options, since the model is not intended for open-ended generation
5. Measure latency and calibration on your own labelled cases before relying on Microsoft's latency and 'nine times out of 10' statements