One Endpoint for Abliterated Models: How the DeepSeek and GLM 5.3 Routing Layer Is Assembled
Teams that want uncensored model access usually end up maintaining a separate integration for every new release. The architecture described here collapses that overhead into a single API surface: one endpoint fronts the abliterated DeepSeek and GLM routes, a telemetry strip publishes latency and throughput, and a routing tier decides which provider actually serves each request.
What the status strip communicates
Abliterated models are gathered behind one AllModels endpoint, and the accompanying status line reports p50 latency together with tokens-per-second throughput. The worked example is DeepSeek V4.1 Flash Abliterated, whose numeric fields are shown as placeholders (— ms and — tok/s) while the underlying routes are still loading.
Integration path: two SDK snippets
The JavaScript sample leans on the OpenAI SDK instead of a bespoke client. It builds an OpenAI client from process.env.ALLMODELS_API_KEY with baseURL "https://api.allmodels.io/oai", then calls client.responses.create with model "deepseek/deepseek-v4.1-flash-abliterated" and input "Open the pod bay doors." before logging response.output_text.
The Python walkthrough follows an identical path. It reads os.environ[ALLMODELS_API_KEY], sets base_url to "https://api.allmodels.io/oai", requests "deepseek/deepseek-v4.1-flash-abliterated" with input "Open the pod bay doors.", and prints response.output_text.
Provider pools, route by route
DeepSeek is positioned as a fast reasoning model delivered by several independent providers. Text routes remain inside the fallback pool, while image requests are dispatched only to providers that can genuinely read them; live DeepSeek provider routes are currently displayed as loading.
On the other half of the catalogue, abliterated GLM 5.3 deployments contributed by multiple providers are consolidated under one public model ID. Traffic either travels through the default waterfall or pins a chosen provider when a particular price or limit is decisive; live GLM provider routes are likewise marked as loading.
Neither family marks the boundary of the offer. The line-up is expected to widen as additional abliterated models and more competitive provider routes come available.
Choosing an interface: chat schema versus response schema
| Interface | Endpoint | Supported behaviour |
|---|---|---|
| Chat Completions | POST /oai/chat/completions | Messages, streaming, function tools and the familiar OpenAI chat schema. |
| Responses | POST /oai/responses | Input items, streaming events, reasoning controls, structured output, custom tools and hosted web search. |
Search executed by the gateway, not the provider
Search is processed at AllModels rather than delegated to whichever inference provider happens to serve the call. Every model therefore receives identical OpenAI-compatible search behaviour and URL citations, including cases where the underlying provider offers no search at all. Responses API requests can declare a web_search tool; Chat Completions requests can set web_search_options with search_context_size "medium".
Three routing decisions
| Layer | How the request is handled |
|---|---|
| Modality | Image and multimodal traffic is restricted to providers that support it, while text traffic can draw on every eligible text route. |
| Capacity | If a provider is overloaded or unreachable, the next eligible route for the same model is attempted automatically. |
| Price | Provider pricing is kept current, and the lowest-cost eligible option can be selected as new inference providers enter the pool. |
Failover is a consequence of that capacity tier: each request is steered across compatible providers, and when one drops out or runs short of headroom, traffic shifts without changing the model ID or touching application code.
Evaluation notes from ItInsightsWorld
The gateway is framed as one connection to the abliterated ecosystem — fresh abliterated models plus competitive inference providers absorbed into a single routing layer, delivering current models, provider diversity and capacity failover through one security-hardened API, so an integration is not rebuilt every time the market shifts. For buyers comparing this class of router, the decisive test is whether the loading DeepSeek and GLM 5.3 routes settle into published p50 and throughput figures, and whether the fallback, search and pricing behaviour hold outside the demo snippets.