Five inference providers now serve abliterated DeepSeek models behind one endpoint

A model-routing service called AllModels now has five inference providers serving abliterated DeepSeek models behind a single API endpoint, according to a post that reached the Hacker News front page and pointed to the company's abliterated model router page.

The setup means developers can request one model identifier, such as deepseek/deepseek-v4.1-flash-abliterated, and have the request fulfilled by whichever of the five participating providers is available, without changing application code.

What abliteration means

Abliteration is a post-training technique used on open-weight language models to remove their built-in refusal behaviour. Instead of fine-tuning on new data, practitioners identify the direction in the model's activations that corresponds to refusing a request and neutralise it. The result is a model that answers prompts the original would have declined. Because the weights are open, the resulting models can be run by anyone, and, as this story illustrates, hosted commercially by inference providers.

How the router works

According to AllModels, the service exposes an OpenAI-compatible API, so existing integrations can switch over by changing the base URL and API key in the standard OpenAI SDK. Two endpoints are supported: chat completions at /oai/chat/completions, covering messages, streaming and function tools, and a responses endpoint at /oai/responses, covering input items, streaming events, reasoning controls, structured output, custom tools and hosted web search.

The company describes several routing behaviours on the page:

Automated fallback. Each request is routed across compatible providers, and if one is unavailable or capacity-constrained, traffic falls back to the next eligible route. The model ID and application code stay the same.

Modality-aware routing. Image and multimodal requests are sent only to providers that can actually handle them, while text requests can use every eligible text route.

Price-based routing. AllModels says it keeps provider pricing current and can route to the lowest-cost eligible option as new providers come online.

Router-hosted web search. Search is processed at AllModels rather than delegated to whichever inference provider serves the request, which the company says gives every model the same OpenAI-compatible search calls and URL citations, even when its provider does not offer search natively.

Beyond DeepSeek

The DeepSeek models are not the only ones covered. The page also lists abliterated GLM 5.3 deployments from multiple providers sharing one public model ID, with the option to pin a specific provider explicitly when a particular price or rate limit matters. AllModels positions DeepSeek and GLM as the starting point for a broader catalogue that will grow as new abliterated models and provider routes come online.

The listed DeepSeek route is described as a fast reasoning model served by several independent providers.

Why it matters

Five providers behind one identifier is notable for a class of models that mainstream hosted APIs typically will not carry. It removes single-vendor dependence for anyone running abliterated models in production: capacity problems at one provider no longer take an application offline, and price competition plays out at the routing layer rather than requiring the developer to re-shop for hosting.

The OpenAI-compatible surface also lowers switching costs to near zero, since adopting the router is a configuration change rather than a rewrite. Features that were previously provider-specific, such as fallback, cheapest-route selection and web search with citations, are being pushed up into the routing layer where every underlying model inherits them.

The flip side is access. Making uncensored open-weight models as easy to consume as any mainstream API broadens their reach for legitimate uses such as research, red-teaming, fiction writing and work in regions with heavy content filtering, but it equally reduces friction for misuse. The continued debate over abliterated models is largely about that trade-off, and services like this one make the access side of it materially easier.