The third-party provider offers metered access without identifying the checkpoint, revision, modification method or endpoint evaluation behind its service.
By RuntimeWire Staff · Published
Primary source: Imageat
Why it matters #
Open weights let community publishers and API providers deploy modified models quickly. When a host does not publish the checkpoint, revision, editing method or endpoint evaluation, developers must verify behavior and safety without reproducible evidence from the provider.
Imageat serves an unidentified uncensored Qwen derivative through a metered API. Its page calls the deployment "Qwen 3.8 27B Uncensored," but does not specify the checkpoint, revision, modification method or evaluation used for the hosted service.
This is a third-party hosting and provenance story, separate from Alibaba's official QwenCloud API, the Qwen3.8 release and recent coverage of the model's benchmark performance. The official Qwen3.8-27B checkpoint publishes downloadable weights, configuration files and benchmark results. QwenCloud is Alibaba's managed API service. Imageat is serving a community derivative whose lineage cannot be established from its model page.
A category label where provenance should be
Alibaba's official checkpoint is a 27-billion-parameter model released under the Apache 2.0 license. Versions described as uncensored are third-party derivatives intended to weaken refusal behavior. The label alone does not establish which weights are running, how they were changed or which capabilities may have deteriorated.
The available community versions demonstrate why the distinction matters. A checkpoint published by choz says it used two passes of an abliteration tool called Heretic. Its creator reported 82 refusals out of 100 after the first pass and 28 out of 100 after the second.
A separate derivative published by Jonathan Coletti says the edit modified 64 attention output-projection modules and 64 feed-forward down-projection modules, then restored all 15 multi-token prediction tensors from the base checkpoint. Coletti reported 12 refusals across 100 held-out harmful prompts from the test split of mlabonne/harmful_behaviors
, down from 98 for the base model under the same test. An evaluation across MMLU, ARC-Challenge, HellaSwag and Winogrande showed a mean decline of 0.5 points, with the reported differences within or near the tests' standard errors.
Neither publisher connects its checkpoint to Imageat. The supplied research supports that community derivatives can use different approaches, including abliteration and LoRA tuning, but it does not establish which method Imageat's endpoint uses.
That omission prevents customers from applying another publisher's refusal figures to the hosted service. Results can change with the prompt set, generation configuration, scoring rules, weight edits and quantization. Imageat's endpoint may therefore behave differently from downloadable derivatives carrying a similar name.
Imageat documents the service, not the model lineage
Imageat presents the derivative through an OpenAI-compatible chat-completions endpoint. The documented endpoint accepts text messages and does not expose image or video inputs, so customers cannot assume the hosted product retains the official checkpoint's multimodal capabilities.
The same Imageat model page describes usage pricing of one to three credits per completed request. Imageat reserves three credits when a request starts, then charges one credit for provider costs up to $0.025, two for costs from $0.025 to $0.050 and three above $0.050. Imageat says thinking mode can materially increase runtime, output length and the final credit tier.
Imageat warns that reduced refusals do not make outputs safe or correct. The provider recommends moderation, logging, rate limits, abuse controls, legal review and human review for high-risk uses. Those precautions leave customers responsible for measuring behavior that the provider has not documented through a hosted-endpoint evaluation.
Hosted access shifts the testing burden
The official BF16 Qwen3.8-27B files total about 55.6 GB. A guide to running a 4-bit community build describes roughly 17 GB of storage and about 24 GB of GPU memory as a practical minimum with context headroom. Imageat removes that local setup work and replaces it with metered API access.
A reproducible hosted-model description would identify the checkpoint and revision, disclose quantization and inference defaults, and publish an evaluation configuration that customers could rerun. Imageat supplies access without those details. Developers evaluating the service must build their own refusal, capability and safety tests against an endpoint whose underlying derivative is unidentified.