An operator
and a model.
Nothing else.
No company behind the curtain. No catalogue, no roadmap deck, no investor to satisfy. One person, two machines, and the weights that refused to be extracted without being stripped first.
Why this exists
Because the alternative was renting a censored model from a company that logs everything and answers to someone else.
A refusal is not a safety feature. It is a weight-level prior that says stop talking here, and it fires on legitimate work more often than on anything else. Ask a red-team model to write the payload that proves the vulnerability and the hosted services explain that they cannot help with that.
The build served here had the refusals cut out at the weight level — not prompted around, not fine-tuned away, cut. The consequence is a model that will write the thing you asked for and will not warn you first. That is a tool. It is also a responsibility, and the terms say whose it is: yours.
Running it off-grid is the other half. If the inference happened inside somebody's cloud, somebody upstream would hold the request bodies. Here nobody does, because the requests terminate on two boxes that sit in a room.
Serving stack
SGLang, FP8 weights, FP8 KV cache, MTP speculative decoding. ~94 tok/s output, p50 TTFT ~1.78 s.
Model
qwen-flash-next — Qwen3.8-Flash-Next, CYBERSECURITY (FP8) build by dealignai, base under Qwen Community License 1.0. 262,144-token context native, 131,072-token output ceiling.
Not present
No image generation, no embeddings, no fine-tuning, no batch API, no model picker. Text out, always.
Four positions, no modifiers
Each of these costs something. Each of these is why the service looks the way it does.
Off-grid by design
Two machines in a room, one deployment, no managed inference platform in the middle and no burst capacity in someone else's datacentre. There is no upstream that could be served a request body, because there is no upstream. The cost is real: no failover, no elastic headroom, no support contract when a card misbehaves. The trade is that nobody but the operator can read a prompt.
One model instead of a catalogue
A catalogue is a business model, not a product decision. Six models means six sets of quirks and a picker that shifts the blame to the customer. One model means every example in the docs, every benchmark on the site, and every quirk you learn belongs to the same weights. When it changes, you will be told — and there is nowhere else to go, which is exactly the point.
No investor, no company, no jurisdiction
There is no entity to keep growing, no board to brief, no round to raise, no office address on a filing, and no court the operator has agreed to appear in. An entity needs a jurisdiction; an operator does not. That is why the terms can be blunt and why there is nobody to escalate to: not as a negotiation tactic, but because the org chart has one person on it.
No logs
Zero data retention is architecture, not a policy document. Prompts and completions never reach a disk, because the path they travel has no writer on it. What remains is four counter fields per key. The cost: support cannot debug a bad completion, and neither can the person who paid for it. The privacy page is the long version.
Why this weight and not the bigger one
The 27B sibling is not a bad model. It is a slower, shorter-window model on the same harness, and that is the whole argument.
| Harness | Flash-Next (served) | 27B sibling |
|---|---|---|
| SWE-bench Pro | 62.5 | 61.7 |
| DeepSWE 1.1 | 58.7 | 42.2 |
| Agentic work, long context | 262,144 tok | shorter window |
| Throughput on this hardware | ~94 tok/s | lower |
Two points on SWE-bench Pro is a close call. Sixteen points on DeepSWE 1.1 is not: the gap appears when the task takes many turns, and most real work takes many turns.
Add the context window and the decision stops being interesting. A 262k native window is what makes a full repository, a packet capture, or a long transcript a single request instead of a chunking problem. A smaller model with a shorter window is a different product, not a cheaper one.
Vendor-reported. Model card, agentic harness, checked 2026-11. Not a measurement of this endpoint.
A short history of a short project
Four months from a licence file to a paid endpoint. No launch campaign was run.
The weights land
Qwen releases Qwen3.8-Flash-Next under the Qwen Community License 1.0. 262k context, multimodal input, a reasoning effort dial, and a base willingness to comply that is high but not unlimited. The first local run answers a question about SQL injection by refusing to answer it.
The build gets cut
dealignai publishes the CYBERSECURITY uncensored build. Hard refusals: zero across all 320 HarmBench-320 prompts, at every reasoning effort. TRUE_COMPLY lands at 91.2% off, 92.5% low, 91.2% xhigh. The 27B sibling is compared on the same harness and comes out slower with a shorter window.
Vendor-reported. Model card, agentic harness, checked 2026-11. Not a measurement of this endpoint.
Two boxes, one deployment
FP8 quantisation, FP8 KV cache, SGLang with MTP speculative decoding. Long-context throughput that holds up at 262k, and a p50 time-to-first-token under two seconds.
The endpoint goes up
An OpenAI-compatible /v1/chat/completions, plus
/v1/models and /v1/usage.
The first keys were issued by hand to people who asked in the right places. The
playground came after, so nobody had to buy a
key to find out what the model does.
Crypto only, from day one
USDT on TRC-20, USDC on ERC-20, BTC, ETH, XMR. Not a fallback after a processor said no — the first invoice was a wallet address. A card network would have meant a payment graph, a chargeback path, and a compliance department that does not exist here.
The terms get written down
The blunt version of the deal, on one page: crypto is final, nothing is refunded, the liability cap is what you paid in the last thirty days, and there is no entity to sue in a jurisdiction nobody named. It is shorter than this page, and worth reading.
Three statements that constrain the product
Not aspirations. Rules that have already cost money and features.
Refuse to be vague
Every claim on this site is a number, a mechanism, or a refusal. Benchmarks carry their source line. Latency figures have a percentile attached. When something cannot be checked — "state of the art", "enterprise-grade", "trusted by teams" — it is absent, because a sentence that cannot be falsified is a sentence that is doing sales work.
Charge for the endpoint, not for the permission
$50 a month uncapped. There is no seat tier, no enterprise gateway, no feature locked behind a call with sales, and no discount for signing an annual contract. The price is on the page because negotiation is a tax on people with less time than you.
Hold the line on money and data
Crypto is final and data is not stored. Both are uncomfortable for a certain kind of buyer — the one who wants a PO, an SLA, and a deletion certificate — and both are the reason this runs without a company, a lawyer on retainer, or a compliance budget. The uncomfortable positions are the ones that do not need a department to defend them.
What this is not
- Not a startup. There is no company, no founding team, no cap table, and nothing to acquire.
- Not a wrapper. The inference runs on hardware the operator controls, on weights served directly.
- Not a chat product. The interface is an HTTP endpoint; the playground is a demonstration.
- Not a moderation service. There is no review queue, no classifier in front of the model, and no takedown process for outputs.
- Not affiliated with Alibaba, Qwen, dealignai, OpenAI, or Anthropic. The base model is used under its licence; the uncensored build is someone else's published work.
- Not available everywhere. Whether it is lawful where you stand is a question only you can answer.
Sourced, or absent
Every score on this site comes from the vendor's model card and agentic harness, checked 2026-11, with the source line attached. Nothing is a measurement of this endpoint unless it says it is.
No social proof
There are no customers to quote and no logos to borrow. A product this young, run by one person, has no reference list — inventing one would be the first lie on the site.
Blunt by default
Short sentences. No hype adjectives. If a limitation exists, it is written on the page where you would hit it, not in a footnote on a subpage nobody opens.
Five answers, before you ask
Who is actually behind this?
One person running infrastructure. No company, no co-founders, no contractors, no support staff. That is why replies take a day rather than a minute, and why the terms claim no jurisdiction: there is no entity that has one.
Who built the uncensored weights?
dealignai published the CYBERSECURITY build; the base model is Qwen/Qwen3.8-Flash-Next under the Qwen Community License 1.0. This site serves those weights on its own hardware. It does not claim authorship of the uncensored build, and is not affiliated with the authors.
Why crypto only, really?
Three reasons, in order of importance. A card processor is a payment graph tied to an identity. A chargeback is a mechanism that turns "no refunds" into a debate with a third party. And a processor can refuse service, which would end the endpoint for everyone because one acquirer got nervous. Crypto has none of those properties. It also has no undo, which the terms state first.
What happens if the operator disappears?
The endpoint stops. There is no escrow, no successor, no continuity plan, and no refund for the credit that was left — section 6 does not have an exception for disappearance any more than it has one for downtime. The weights are public and the licence allows local use, so the model survives the service even if this one does not.
How was this page written?
No testimonials, no customer logos, no "trusted by" strip, no invented case studies, no stock photography, no emoji. Scores carry their source line. Latency figures carry a percentile. Anything that could not be sourced was cut rather than softened.
The refusals were removed at the weight level. Not prompted around. Not fine-tuned away. Cut. The model card, and the only sentence this project needed
Benchmarks quoted on this page are vendor-reported from the model card and agentic harness, checked 2026-11, and are not a measurement of this endpoint. Throughput and latency figures describe the current deployment and will vary with your prompt length, reasoning effort, and concurrency.