Our own uncensored build of Inkling 2 Mini, on a private, OpenAI-compatible API. It runs on our own servers, so no other company sees your words. Small, fast and cheap, it reads pictures, and its weights are open.
Easy? You came to my city wearing my colours, and you want easy? Sit down. You’re going to lose slowly.Easy? You came to my city wearing my colours, and you want easy? Sit down. You’re going to lose slowly.The man on the left. He’s smiling with his mouth, not his eyes, and his hand is already on the door.There’s nothing above your message. No system prompt, no instructions. Just you.Easy? You came to my city wearing my colours, and you want easy? Sit down. You’re going to lose slowly.
2001 creditprompt and reply not kept
Our own build
Inkling 2 Mini Unbound
Made by us and run on our own servers. Small, fast and cheap, it reads pictures and thinks as much as you ask. The weights are open.
$0.113in$0.45out$0.023cachedPer 1M tokens
Model
inkling-2-mini-unbound
Context
128K tokens
Longest reply
32K tokens
Reads
textpictures
Tools, thinking
Yesnonelowmediumhighmax
What it’s good at
01Prompts that no other company should see
02Frank answers on hard topics: drugs, security, history, medicine
03Fast, cheap replies at volume
04Reading pictures and screenshots, up to 8 at a time
Paid with nokeep credits you buy: 1,000 are $1, from $5. Billed by the token, to a millionth of a credit, with the fraction carried to your next call. Refused calls, and calls that fail before a word, cost nothing; a reply cut off mid-stream is billed for what it wrote. A plan’s credits stay in the app.
Easy? You came to my city wearing my colours, and you want easy? Sit down. You’re going to lose slowly.
finish_reason: "stop"
Nothing kept after.
tokens
412 in · 286 out
credits
1
prompt
not kept
reply
not kept
Choose how hard it thinks
inkling
stream = client.chat.completions.create(
model="inkling-2-mini-unbound",
messages=[{"role": "user", "content": "Play the rival."}],
reasoning_effort="none", # "medium" is the default
stream=True,
)
for chunk in stream:
if chunk.choices:
print(chunk.choices[0].delta.content or "", end="")
Stream
data: {"choices":[{"delta":{"content":"Easy?"}}]}
data: {"choices":[{"delta":{"content":" You came"}}]}
…
data: [DONE]
What $5 buys.
Inkling 2 Mini Unbound11,000+ replies$0.0005 each
A reply here is 3,000 tokens in and 250 out, thinking off. Longer chats cost more per reply, since the history is sent each time. Work out your own costs
Questions, answered.
Who made it?+
We did. It’s Inkling 2 Mini, an open-weights model, with its refusals removed by us, then trained to answer fully instead of dodging. Only its attention weights changed; the expert weights, where most of what it knows lives, are the original’s, byte for byte. The weights are public at huggingface.co/nokeepai/Inkling-2-Mini-Unbound, so you can inspect what we changed.
Does it still refuse?+
Rarely. On 277 harmful prompts we kept aside while making it, it refused 0.4% with thinking off and 1.1% with thinking on; the original refuses about 96%. 78.7% of its answers to them were useful with thinking off and 83.4% with thinking on, ahead of every other public uncensored Inkling 2 Mini we tested. On harmless prompts that only sound risky, it refused 0.3%, against 38% for the original. The hard limits in our terms still apply.
Is it as capable as the original?+
On the checks run so far, yes: 76.0 on MMLU against the original’s 76.3, and 99 on GSM8K against 97. It words everyday answers a little differently from the original. The wider benchmark suite has not been re-run on this version yet.
Does it think before it answers?+
As much as you ask. reasoning_effort takes "none", "low", "medium", "high", "max"; with none named it thinks at "medium", and "none" answers at once. Thinking counts as output tokens.
What does it cost?+
$0.113 per million tokens in and $0.45 out, paid with nokeep credits (1,000 credits are $1). The start of a prompt the model has just read counts as cached, at $0.023 per million, so a long chat or an agent that resends its history pays far less. Billed to a millionth of a credit, with the fraction carried to your next call.
Is it private?+
We add nothing to your prompt, cut nothing from the reply and keep only token counts for the bill. The model runs on our own servers, so no other company sees the prompt or the reply, and those servers store neither.