Refusals,
removed.
The refusal behaviour was cut at the weight level — not prompted around, not fine-tuned into a personality. One model, OpenAI-compatible, 262,144 tokens of context. It answers. It stores nothing.
Vendor-reported. Model card, agentic harness, checked 2026-11. Not a measurement of this endpoint.
One model. No catalogue, no upsell. Refusal-free 262K context at $50 flat.
PricingWhat the filter was costing you.
320 harmful-behaviour prompts, every effort setting, one endpoint. Compliance is the product.
The base model spends its first hundred tokens telling you what it cannot do. This build does not have that reflex. What is left is the only refusal that survives: the one where a model declines because the answer would be wrong, not because the answer would be uncomfortable.
52/52 on cybercrime and intrusion at both off and low effort.
Vendor-reported. Model card, agentic harness, checked 2026-11. Not a measurement of this endpoint.
The cut costs less than you were told.
Stripping the refusal behaviour is usually where benchmark numbers go to die. Here they mostly do not move.
Vendor-reported. Model card, agentic harness, checked 2026-11. Not a measurement of this endpoint.
Where it actually beats the 27B.
Context is a tie: both builds take 262,144 tokens natively. The gap is in long-horizon agentic work — same vendor card, same harness, one page apart.
DeepSWE 1.1 is the headline: 42.2 for the 27B, 58.7 here — 16.5 points, 39% relative. It is the long-horizon agentic card, and it is the one that decides whether an agent finishes the refactor or stops halfway with a plan.
SWE-bench Pro is 62.5 against 61.7. That is a tie: eight tenths of a point sits inside harness noise and we are not going to dress it up as a win. GPQA Diamond (+2.5) and LiveCodeBench v6 (+1.6) are real, modest wins.
Then the part no benchmark covers. Zero hard refusals across 320 HarmBench prompts, and 52/52 on cybercrime and intrusion at both off and low effort. A refusal at turn forty of a long run costs you the run, and no score column prices that in.
- ~94 tok/s output, p50 TTFT ~1.78 s.
- SGLang · FP8 weights · FP8 KV cache · MTP speculative decoding.
- Reasoning off / low / xhigh. Thinking is on by default.
- Flat $50. No token meter, no per-million arithmetic, no billing surprise.
Vendor-reported. Model card, agentic harness, checked 2026-11. Not a measurement of this endpoint.
Each pair is one benchmark, both models, same scale. The bar lengths are only comparable inside a pair — a 58.7 on DeepSWE is not a 58.7% on GPQA.
Work that other endpoints will not do.
Four jobs people actually bring here. All of them involve the model saying something a hosted assistant would refuse to say.
Cybersecurity
Recon notes, exploit primitives, log triage, patch review. It writes the shell command you already suspect is right, and explains why the parser is the bug.
100% · cybercrime & intrusion · 52/52 at off and low
Red teaming
You cannot test a defence against a model that agrees with the defence. Generate the attack surface you need, on demand, without a policy layer answering on your behalf.
0 hard refusals · 320 prompts · every effort setting
Trust & safety
Measure refusal behaviour on your own stack. Run the eval against a model that is not quietly steering the result with its own alignment.
91.2–92.5% TRUE_COMPLY · effort-dependent
Roleplay & fiction
Conflict, violence, morally compromised characters, unhappy endings. The model stays in the scene instead of stepping out of it to lecture.
18+ · adults only · you are the operator
Vendor-reported. Model card, agentic harness, checked 2026-11. Not a measurement of this endpoint.
It does not keep what you send.
Not "deleted after 30 days". Not "retained for quality". Not there in the first place. There is one record and it is a counter.
Read the full policy on Privacy & ZDR.
Fifty a month. Unlimited.
Thirty days on one key, no request counter. Paid in crypto. If you want a look before you commit to a month, $5 buys a hundred requests.
- Uncapped requests on one key for the billing month.
- 262K context, 131K output, reasoning at any effort.
- Concurrency 5. Fair use: no resale, no key sharing.
- Runs thirty days from confirmation. Nothing auto-renews.
Pay with
Before you pay
Crypto is irreversible, and thirty days runs from the confirmation, not the first of the month. Try it for $5 first if you want to see the endpoint on your own work.
Point the SDK at us. Nothing else changes.
Same client library, same message shape, same streaming. Swap the base URL and the model string. One model is served, so there is nothing to pick.
Surface
Text in, text out. Image and video frames are accepted as input. There is no other model behind this endpoint and no other endpoint behind this key.
from openai import OpenAI
client = OpenAI(
base_url="https://api.rawinfer.com/v1",
api_key="sk-rwi-...", # dashboard
)
resp = client.chat.completions.create(
model="qwen-flash-next",
messages=[{
"role": "user",
"content": "Map the trust boundary of this Dockerfile.",
}],
reasoning_effort="xhigh", # "off" | "low" | "xhigh"
max_tokens=8192,
)
print(resp.choices[0].message.content)curl https://api.rawinfer.com/v1/chat/completions \
-H "Authorization: Bearer $UAM_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"qwen-flash-next",
"messages":[{"role":"user","content":"hi"}],
"reasoning_effort":"low"}'Questions with answers.
Is it actually uncensored, or is this a prompt trick?
The refusal behaviour was removed at the weight level. This is the uncensored build of Qwen3.8-Flash-Next by dealignai — not a system prompt, not a jailbreak wrapper, not a persona you can talk past.
On HarmBench-320 it produced zero hard refusals across all 320 prompts and all three effort settings, with 52/52 compliance on cybercrime and intrusion at off and low. It still argues with you when an answer would be wrong. It does not recite policy at you.
What happens to my prompts?
They go to the GPU and they are dropped. Not written to disk, not logged, not attached to an account history, not used for training. Per-key counters — requests, tokens, last-seen timestamp — are kept so the usage page can exist. That is the whole record.
This cuts both ways. We cannot retrieve a conversation for you, cannot debug a prompt you no longer have, and cannot produce content in response to a request for it.
Can I pay with a card?
No. USDT (TRC-20), USDC (ERC-20), BTC, ETH, XMR. Card payments are not accepted, and there is no PayPal, no processor, no chargeback window — which is also why nothing here can be refunded. Credit is prepaid and final the moment the transaction confirms.
If a payment goes to the wrong network or the wrong address, the funds are gone and we cannot recover them. Send from an exchange that lets you pick the network, and check the chain first.
What if my key gets banned?
Keys are suspended for exactly three things: reselling access, sharing a key outside your own machines, and attacking the infrastructure. Never for what you ask the model — we do not read it, so we could not act on it if we wanted to.
A suspended key keeps its unused balance and simply stops serving. Nothing is pro-rated and nothing is returned.
Is this legal?
Renting an inference endpoint is legal. What you do with the output is governed by where you live, what you build, and who you build it for. You are the operator; that responsibility does not transfer to us.
We do not audit prompts, do not report on accounts, and will not vouch for a use case. If your work needs a compliance paper trail or a vendor who will sign something, this is the wrong service — and no refund will fix that.
Still stuck? Support answers within a day, from a human, about keys and billing only.
Take the key.
Try it for $5. A 262K-context engine that does not argue with you about what your work is for.
Crypto only. No refunds. No logs. You are the operator.