Skip to main content

Managed models and modalities

Check the configured list​

import json
from envloop import Client

with Client.remote() as client:
    print(client.training_jobs.models())
    request = json.loads(open("REQUEST.json").read())
    job = client.training_jobs.submit(request)
    print(job["job_id"])

GET /v1/job/training-models is project scoped and uses the same configured recipe routes as submission. It returns exact base_model, model_revision, recipe policy/hash, modalities and unsupported workloads, plus features (labels, and for Qwen/Qwen3.8-27B the parent checkpoint rules, hyperparameter ranges and loss mask) and limits (sequence length) that submission enforces. readiness=configured is distinct from accepted submission and actual runnable GPU capacity. A 503 means discovery is unavailable; do not infer support from serving profiles or a private backend model list.

GLM-5.3-Flash​

The exact identifier is zai-org/GLM-5.3-Flash; revision is eb9eb208eb0d988989d07a6a12d0fdeb5f52574a. In cloud mode it is a default managed SFT offering for new and existing projects without GLM-specific owner approval. Normal permissions, enforce billing, budget and capacity constraints remain. An emergency deny can remove the offering.

Other recipes​

Qwen and DeepSeek batch recipes depend on current platform configuration. Inspect discovery before submitting; this change does not promise their availability or parity with an interactive SDK.

Images​

See data for bounded inline inputs. The recipe chooses model/GPU implementation; do not send backend/provider/executor selectors.

recipe=null means omit recipe and use the model default; SFT is a workload name, not a submit value for this field. Hello World uses explicit recipe="hello-world" and its fixed four records.

The native managed backend discovers text SFT from upstream capabilities. Its revisions are provider_managed with model_revision null; this does not prove a frozen GLM QLoRA recipe. Native submission rejects GLM without that recipe while retaining the Qwen managed controller.