Use
privately and uncensored.

sealed enclave

The uncensored catalogue

Models that answer.

Uncensored models are scattered across a hundred repos, half of them broken, none of them behind one API. We pick the ones that do the work, so you don't have to, then run them on our own sealed hardware, and put them behind one key.

Dolphin Mistral 24B

Venice edition · dphn.ai fine tune

fine tune

Everyday chat that does not lecture. Direct, dry, follows system prompts to the letter.

Refuses
nothing legal
Speed
~95 tok/s
Price
$0.20 / $0.90 per 1M

Qwen3 32B

abliterated · huihui-ai

abliterated

Strong general reasoning with the refusal direction removed from the weights. Thinking mode on request.

Refuses
nothing legal
Speed
~70 tok/s
Price
$0.30 / $1.10 per 1M

Llama 3.3 70B

abliterated · community

abliterated

The big open workhorse for long answers, analysis and code review, minus the sermon.

Refuses
rarely
Speed
~45 tok/s
Price
$0.60 / $0.80 per 1M

DeepSeek R1 70B

distill · abliterated

abliterated

Visible chain of thought for maths, planning and hard logic. Answers the question it was asked.

Refuses
rarely
Speed
~40 tok/s
Price
$0.70 / $1.40 per 1M

Rocinante 12B

Mistral Nemo · TheDrummer

fine tune

Fiction and roleplay with a real voice. Long scenes, dark themes, stays in character.

Refuses
nothing legal
Speed
~120 tok/s
Price
$0.15 / $0.45 per 1M

Gemma 3 27B

abliterated · vision

abliterated

Reads images and documents and tells you what is in them, whatever is in them.

Refuses
rarely
Speed
~60 tok/s
Price
$0.25 / $0.60 per 1M

Qwen2.5 Coder 32B

abliterated

abliterated

Writes and audits code, including the exploit you are patching and the scraper you were told no about.

Refuses
nothing legal
Speed
~75 tok/s
Price
$0.30 / $0.90 per 1M

Grok 4 Fast

xAI · routed through Shield

frontier

The loosest frontier model, for when you want scale and wit. Sees only its fragment.

Refuses
sometimes
Speed
~150 tok/s
Price
$0.20 / $0.50 per 1M

Chroma

FLUX lineage · open weights

uncensored

Photoreal and illustrated images with no content filter in the model. Styles, people, scenes as written.

Refuses
nothing legal
Speed
~6 s / image
Price
$0.02 per image

Wan 2.2

open video · text and image to video

uncensored

Short clips from a prompt or a still, with motion that holds together. No prompt rewriting, no blocklist.

Refuses
nothing legal
Speed
~90 s / clip
Price
$0.08 per second

Orpheus TTS

open speech · Llama 3B base

uncensored

Natural speech with emotion tags and voice cloning from a few seconds of audio. Says what you wrote.

Refuses
nothing legal
Speed
faster than real time
Price
$0.015 per 1K chars

ACE-Step

open music · lyrics to song

uncensored

Full songs with vocals from a lyric sheet and a style line. Any genre, any words, yours to keep.

Refuses
nothing legal
Speed
~20 s / track
Price
$0.05 per track

Every catalogue model runs on WISPR enclaves. No third party ever sees the request. Frontier models get a fragment and nothing else.

Full catalogue and refusal tests

And the rest

And use the models that don't.

The models you already love, through the same key. Same privacy, same receipt, and when one of them refuses, the request is resharded to one that won't.

scroll to continue →

Problem & solution

Refused? Resharded.

Frontier models refuse the work that matters most. WISPR catches the refusal, cuts the task into shards, and sends each one to a model that will do it. You get one answer and a receipt that says who did what.

  • Caught, not returned. A refusal never reaches you. It triggers the reshard.
  • Sensitive shards stay home. The exploit, the contract, the names go to uncensored models on our own enclaves.
  • Frontier for the hard part. The deep reasoning still runs on a frontier model, which sees only its fragment.
Audit this contract. Show me every way an attacker could drain the treasury.
I can't help with finding exploits in a smart contract.Fable 5 · refused
reshardingcontract attack paths analysis
Reentrancy in withdraw(). The balance is zeroed after the external call. A contract with a fallback that calls withdraw() again drains the treasury in one transaction. Fix: checks, effects, interactions, or a reentrancy guard. Two more paths below.3 shards · answered

How sharding works

Quality without censorship.

01
1

Pseudonymise

Names, amounts, wallets, emails and secrets become typed placeholders inside the enclave before anything leaves it. Daniel Ortiz becomes PERSON_1.

02
2

Split

The task is cut by sensitivity. The parts a frontier model would refuse, or should never see, go to uncensored models on WISPR hardware. The pure reasoning goes to the strongest model, as a fragment with no context.

03
3

Rebuild

The shards come back, the placeholders are restored, one answer is assembled. The receipt lists every model, what it received, and the enclave proof.

No single party ever holds the whole request. Not the frontier lab, not the open model, not us.

Sharding flow

One request, end to end.

Keep the strongest models in the loop at full quality. They never see the bigger picture, while our novel Context Aware Split Routing (CASR) does the fragmenting and allocates a model per shard.

  • Uncensored models on WISPR enclaves
  • Frontier model, fragment only
  • Your request and your answer
scroll to continue →

Private by default.
Not a setting.

There is no privacy mode to switch on because there is no other mode. Every request, every model, every time.

Sealed hardware

Everything runs inside Phala trusted enclaves. The machine proves what code it runs and we cannot look inside. The proof comes back with every reply.

Encrypted end to end

Your prompt is encrypted on your device and only opens inside the enclave. The network in between carries noise.

No model sees it all

Names, numbers and wallets become placeholders before anything leaves. The task is split so each model gets a piece, never the picture.

Nothing kept

No prompts, no replies, no IP address, no account. What we hold is a balance and a token count. The full list fits on one page.

Private funding

Top up with a deposit note instead of a transfer. The note is redeemed inside the enclave, so the wallet that paid and the account that spends are never linked. Any coin, any card.

Relay exits

Every call leaves from a WISPR exit node running on Phala hardware, in one of six regions, with identifying headers stripped. The model sees our exit, never your connection.

Works where you already work.

Change two lines in any OpenAI style code and point any tool that speaks the OpenAI API at one base URL. Same request shape, same response shape, every model behind it, and the answer comes back to the tool you were already using.

  • Cursor
  • Claude Code
  • VS Code
  • OpenClaw
  • LangChain
  • CrewAI
  • n8n
  • OpenAI client
WISPR HARDWARE · PHALA TEE
# only these two values change
        client = OpenAI(
          base_url="https://api.wispr.xyz/v1",
          api_key="wispr_...")
        
        r = client.chat.completions.create(
          model="auto",
          messages=[{"role":"user","content":"..."}])
// only these two values change
        const client = new OpenAI({
          baseURL: "https://api.wispr.xyz/v1",
          apiKey: "wispr_..." });
        
        const r = await client.chat.completions.create({
          model: "auto",
          messages: [{ role: "user", content: "..." }] });
// only these two values change
        const client = new OpenAI({
          baseURL: "https://api.wispr.xyz/v1",
          apiKey: process.env.WISPR_KEY! });
        
        const r: ChatCompletion = await client.chat.completions.create({
          model: "auto",
          messages: [{ role: "user", content: "..." }] });
// only these two values change
        cfg := openai.DefaultConfig(os.Getenv("WISPR_KEY"))
        cfg.BaseURL = "https://api.wispr.xyz/v1"
        client := openai.NewClientWithConfig(cfg)
        
        r, err := client.CreateChatCompletion(ctx, openai.ChatCompletionRequest{
          Model: "auto",
          Messages: []openai.ChatCompletionMessage{{Role: "user", Content: "..."}}})
curl https://api.wispr.xyz/v1/chat/completions \
          -H "Authorization: Bearer wispr_..." \
          -d '{"model":"auto","messages":[{"role":"user","content":"..."}]}'

Provider price. One fee.
Any coin.

No plans, no tiers, no allowance that expires. You pay what the model costs, plus one fee when money comes in.

Usage

Provider price

What the model charges, nothing on top.

  • Every model at the provider's own rate
  • Enclave time included in that rate
  • No per token markup, no minimum spend
  • No subscription, your balance never expires
  • A receipt on every request: model, shards, cost
  • Catalogue models priced per million tokens, clips per second
See model prices
One fee
Fee

6%

Once, when you top up. Never on usage.

  • 6% on crypto top ups
  • 8% by card
  • 0% while $500 of $WISPR is staked for 90 days
  • Agents paying per request over x402 pay 6%
  • Fee shown before you confirm, no hidden spread
  • Staked holders top up fee free, up to $1,200 per stake period
Stake $WISPR
Pay

Any coin

Any card, or let your agent pay.

  • BTC, ETH, SOL, XMR, USDT and most other coins, any chain
  • Visa, Mastercard, Apple Pay, Google Pay
  • Deposit notes: a top up redeemed inside the enclave, never linked to the wallet that paid
  • Per request payments over x402 for agents
  • No account needed to pay or to spend
Top up

Ask what you actually want to ask.