LLM API with no added content filter: Dolphin, Hermes and open models for roleplay and creative work

WeightsAPI’s gateway checks your API key, its model allowlist and your balance, then passes your messages to the model you chose. It adds no content filter of its own: no moderation pass, no topic blocklist. That is the exact claim on our FAQ and trust page.

Three things still shape each answer: the model’s training (Dolphin and Hermes are tuned to reduce refusals, not to guarantee answers), your instructions (system prompt, character card, sampling settings), and the serving configuration, since the chat template must match the deployment.

No added filter does not make anything legal. You remain responsible for lawful use and third-party rights, and each model’s license applies: Apache 2.0 for Dolphin, and the Llama 3.1 terms and use conditions for Hermes 4 70B, a Llama 3.1 70B derivative.

Two catalog models are fine-tunes built for this work; three original releases cover cheaper drafts or a different voice. All run on the same key and balance, so you can switch per character or per scene.

Our pick for most roleplay and long-form fiction is Hermes 4 70B: the larger model, output at $0.40 per million tokens against $0.90 for Dolphin, and positioned by Nous Research for creative writing, roleplay and conversation mixed with reasoning. Choose Dolphin for image input or its character-first tuning, built with Venice. Compare them in Dolphin vs Hermes 4 70B.

The original releases are not refusal-reduced fine-tunes. Use Mistral Small 3 for cheap short scenes within 32K, or Llama 3.3 70B and Qwen3 235B A22B for longer ones. All 15 models are on the models page.

In SillyTavern’s API Connections, choose Chat Completion, then Custom (OpenAI-compatible). Enter https://weightsapi.com/v1 as the endpoint without appending /chat/completions, paste your key without the Bearer prefix, and select NousResearch/Hermes-4-70B or dphn/Dolphin-Mistral-24B-Venice-Edition. Send a Test Message to confirm. Details: SillyTavern integration guide.

Two settings protect your balance. Each request reserves credit for its input plus the maximum output you allow, then settles on actual usage, so a sensible response length keeps credit free in long chats. And give SillyTavern its own key with a lifetime spending cap and a Dolphin and Hermes allowlist, revocable from the console (key controls). The same endpoint suits Open WebUI and LibreChat.

1. Nothing added between your prompt and the model

Your character card and story context reach the model without a moderation layer from us. Fiction with villains, dark themes or hard drama is not blocked by a blanket filter; our line is lawful use and the model’s license. Create a username account.

2. Hermes 4 70B, which the big aggregators don’t list

A 70B creative fine-tune at $0.13 / $0.40. Venice, OpenRouter, NanoGPT and Chutes don’t list the 70B (25 September 2026). OpenRouter carries Hermes 4 only as the 405B at $1.00 / $3.00, plus Hermes 3 70B at $0.70 / $0.70; NanoGPT lists Hermes 4 405B at $0.30 / $1.20 and Hermes 3 70B at $0.408 / $0.408. See the Hermes 4 70B page.

3. Dolphin at Venice’s own price

$0.20 / $0.90, the same as Venice’s API for Venice Uncensored 1.2 and OpenRouter’s listing, which Venice serves. NanoGPT charges $0.40 / $1.80. No premium for Monero or for skipping the email. See the Dolphin page.

4. Monero, Bitcoin or stablecoins, with no platform deposit fee

XMR, BTC, LTC, SOL, ETH, TRX, USDT on TRON or Ethereum, or USDC on Ethereum; the full USD amount becomes credit within about a minute once the network confirmations are reached, and you pay only the network fee. For API credit, Venice takes USD or USDC, OpenRouter USDC with a 5% crypto fee, Chutes TAO or crypto through Stripe. Every asset and network.

5. Username and password, nothing else

No email, phone, ID or card, so your stories are not tied to an inbox, and the gateway does not store prompt or response text, though the inference provider receives each request to answer it. No KYC is not anonymity: usage and payment records stay linked to your account. What we ask and record.

6. A dedicated GPU from the same balance

For a character app with steady traffic, the balance prepays a dedicated GPU: $416 per month for an L40S (48 GB), $1,352 for an H100 SXM (80 GB), no per-token charges on that machine, fit confirmed by quote. Venice, OpenRouter and NanoGPT sell no monthly GPU on pages we reviewed; Chutes bills private GPUs from $1.80 per hour plus a $5.40 deployment fee. See dedicated GPU plans.

Same models, different terms. Prices are USD per million input / output tokens from public pricing pages or model lists.

Content terms differ too: Chutes’ terms prohibit generating “harmful, illegal, or offensive content”, a broad standard for a horror writer. We add no gateway filter and ask for lawful use.

Where others fit better: NanoGPT to start with a few dollars and no account; Venice for its consumer app, privacy modes and separate Role Play Uncensored model ($0.50 / $2.00); OpenRouter for a 460-model catalog and cheaper original models, such as Llama 3.3 70B at $0.10 / $0.32 against our $0.35 / $0.40. More in our OpenRouter alternative comparison.

What will it really cost? Tokens at published rates plus your network fee; no subscription. Roleplay is input-heavy because each turn resends the chat history. At the pricing calculator’s default 80% input / 20% output split, USD 100 buys about 540 million Hermes 4 70B tokens or about 290 million Dolphin tokens (our arithmetic).

Why a USD 20 minimum? Each top-up is a separate on-chain payment with its own network fee and verification; a USD 20 floor keeps that overhead small, and with no platform deposit fee all of it becomes credit. It is per top-up, not a balance to maintain. To spend $5 on a test, NanoGPT fits better.

Can I use it today? Create your account, deposit from $100, credited in about one minute. A published price alone does not mean a model is serving.

How do I pay, and how long does it take? Checkout fixes the crypto amount for 10 minutes, and the transfer must use the listed network; credit lands within about a minute once the network confirmations are reached (BTC 1, LTC 6, XMR 10). Wrong-network transfers have no recovery policy. See billing and crypto deposits.

Will I be asked for ID later? Not in the current account flow. Payment records stay tied to your account, and late or wrong-amount transfers go to review.

Can I get my money back? Credit cannot be withdrawn to a wallet, so top up for planned usage. The only refund is cancelling a GPU order before activation, back to your balance. Charges are grouped by UTC hour, model and key.

Choose WeightsAPI if you write interactive fiction, run SillyTavern characters or build a companion or game-dialogue app, want Hermes 4 70B or Dolphin with no gateway filter, and prefer paying in Monero, Bitcoin or stablecoins to handing over an email and a card.

Look elsewhere if you want to spend under USD 20, need guaranteed capacity or a latency commitment today, or need a withdrawable balance.

Next step: create a username account, store the recovery code offline, make a SillyTavern key with a cap and a Dolphin and Hermes allowlist, and top up once /status shows your model.