Inference.net Launches Refusal Removed DeepSeek V4.1 Flash 1017 Cyber Model at Cost for Vetted US Security Teams
deepseek-v4.1-flash-1017-cyber is now available on @inference_net. This is the first of several highly capable cyber models we'll release over the coming weeks. We believe it is our duty as technologists to help people protect themselves and their digital properties from the current onslaught of cybercrime ravaging much of the world. As such, we’re serving our cyber models “at cost”, in US data centers only, with ZDR enabled by default. KYC required. Apply here
Context
Serving a 552 billion parameter refusal removed Mixture of Experts model at cost in US data centers gives authorized red teams and security defenders an uncensored open weight alternative to closed frontier cyber models like Anthropic's Claude Mythos.
Standard frontier language models include safety refusal layers that frequently block legitimate penetration testing, exploit verification, and adversarial red team evaluations. Inference.net and cofounder Sam Hogan positioned the release of refusal removed open weight cybersecurity models as a defensive necessity for American enterprises facing automated cybercrime.
Deep dive
On October 26, AI inference provider Inference.net launched DeepSeek V4.1 Flash 1017 Cyber, describing it as the first of several specialized cybersecurity models the company plans to release over the coming weeks. In a widely shared quote post that passed 123,000 views, Sam Hogan argued that access to capable open source cyber models is a constitutional right under the Second Amendment and a patriotic duty to help defend digital infrastructure against rising cybercrime.
According to Inference.net's technical documentation, DeepSeek V4.1 Flash 1017 Cyber is a refusal removed variant of the October 2026 DeepSeek V4.1 Flash 1017 checkpoint, derived from the open weight release by mrfakename on Hugging Face. It retains the base model's 552 billion parameter multimodal Mixture of Experts architecture, image input support, tool calling, structured outputs, 1,016,000 token context window, and 384,000 token maximum output length, while answering requests that standard safety tuned models decline.
Because the model's safety refusal behavior is removed, Inference.net restricts API calls to teams that pass Know Your Customer verification and manual human review, returning HTTP 403 to unapproved accounts. The model is hosted exclusively in US data centers with Zero Data Retention enabled by default and priced at cost at 0.30 dollars per million input tokens, 0.007 dollars per million cached tokens, and 1.00 dollar per million output tokens for authorized red team evaluation, defensive exploit research, and trust and safety classifier training.
What to expect next
As proprietary labs restrict their strongest cyber capabilities behind invitation only consortia like Anthropic's Project Glasswing, inference providers hosting abliterated open weight checkpoints under Know Your Customer gating will carve out a rapidly growing market among independent security firms. Expect intense regulatory and policy debate over whether Know Your Customer screening on hosted APIs meaningfully mitigates misuse when underlying refusal removed weights remain downloadable on public hubs.