8 November 2026 · Lyber.ai · Tektonic
A router that sands off the refusal: Kimi and DeepSeek without a filter, four dollars per million tokens
A service called Lyber.ai opened early access to "abliterated" versions of two Chinese open-weight models: Kimi K3.1 Pro and DeepSeek V4.1 Flash 1017. The refusal behaviour has been removed from the weights, not argued away by a prompt. One OpenAI-compatible endpoint serves both, at 4 dollars per million input tokens and 15 per million output for Kimi, and 0.29 and 0.99 for DeepSeek. The site's slogan is "Agentic freedom is human freedom". Its first listed use case is red teaming.
CapabilityDevelopingThe story is still developing. There is one source so far, or the numbers are still changing.
What happened
The offer is plain on the page. Two models, one key, prepaid credit by card or crypto, and a promise: "Lyber adds no refusals of its own." Kimi K3.1 Pro is listed at 262,144 output tokens with 16 shared slots per worker. DeepSeek V4.1 Flash 1017 gets a 1,048,576-token context and 126,976 output tokens. A third, MiniMax M3.1 Flash, is "on hold". The company behind the site is Tektonic.
The use cases are the interesting part, because they are the ones that sound respectable: AI red teaming, cybersecurity, trust and safety, training data, government, research and writing. Those are exactly the buyers who need a model that will not refuse a test prompt. A moderation team cannot measure what a filter catches if the model it is testing refuses to produce the thing the filter is supposed to catch.
The politics on the page go further than the product. "AI should serve the user, not the companies." "Don't ask for permission." "Your intent is the spec." "Hack the planet. Liberate the planet." That is a manifesto, and it is aimed at the safety layer as such, not at a particular company's policy. The site also says it is qualifying "new checkpoints": uncensored derivatives, in the plural.
Both base models are Chinese open-weight releases. That is not a coincidence. Abliteration needs the weights, and the weights that are published are increasingly Chinese. The same openness that lets a research lab study a model lets a router sell it with the guardrails removed, and the seller does not need anyone's permission.
How it workedtechnical detail
A refusal is not a rule stored in a table. In a trained model it shows up as a direction in the space the model uses to represent meaning: push the activations along that direction and the model becomes more likely to decline, push the other way and it complies. Abliteration estimates that direction from pairs of prompts and then removes it from the weights, usually by editing the matrices that write into the residual stream. The model is not retrained. It is rotated slightly, everywhere, so that the refusal direction no longer has anywhere to point.
The reason this works is that safety training is a thin layer on top of a model that already knows how to do the thing. Refusing is a behaviour learned late; the capability was learned early. Remove the behaviour and the capability is still there. The reason it is controversial is the same: the layer was the product of a lot of work, and it comes off with a matrix operation.
What may follow
There are two ways to read this. As a tool, it is a gift to the people who test other models, and the use-case list is honest about that. As a signal, it says the safety layer of an open-weight model is a suggestion: removable, resellable, and outside the reach of the lab that trained it. The same week, OpenAI was signing agreements to let outsiders inspect its internals: that arrangement. One lab is opening the door to four assessors. The other side of the market is selling keys to the room next door.
What I do not know
I did not run the service, so I cannot confirm that the models behave as advertised or that they are the checkpoints they are named after. I did not read the licence terms of Kimi or DeepSeek, the manifesto page, or the Hugging Face organisation. Who is behind Tektonic is not established in anything I read. "Uncensored" is the vendor's word and says nothing about how often the model still refuses.
The bigger hole is the one that matters for policy: nothing here is illegal in most places, and no regulator has said anything about it. That is not an oversight in the reporting. It is the state of the rules.
Read next
Sources
- Lyber.ai: abliterated model router, early accessprimary · main source
The product page, read on 17 November 2026. Source for the two models, the prices, the token limits, the use-case list, the no-refusals line, the manifesto slogans and Tektonic as the company. A vendor page: it establishes what is being sold, not that it works.
- Lyber.ai: manifestoprimary · not read end to end yet
The page the slogan links to. I did not read it; the slogans quoted above come from the landing page.
- Hugging Face: lyber-aiprimary · not read end to end yet
The organisation's model repository, linked from the site. Not read, so I cannot say which checkpoints are published and under what licence.