Eight models. None of them lecture.
Open weights without the refusal reflex. Each model lists how refusals were removed, its base model, context and features. Prices per 1M tokens, metered against a capped lease and charged in gwei at a frozen Chainlink quote; ETH shown at $3,412 (example).
- abliterated
- 2 models
- The refusal direction is projected out of the weights; nothing else is retrained.
- heretic
- 1 model
- Abliterated with Heretic, an automated abliteration tool.
- uncensored
- 5 models
- An uncensored fine-tune of the base model; the exact method is not published.
- context
- up to 1M
- Longer prompts are refused, not truncated.
GLM 4.7 Flash Heretic · GLM
Cheapest. Labelling, extraction, quick drafts.
0.000029 / 0.000152 ETH per 1M at $3,412 (example)
- in / 1M
- $0.10
- out / 1M
- $0.52
- tag
- Small
- base
- GLM-4.7-Flash
- kind
- heretic · Abliterated with Heretic, an automated abliteration tool.
- family
- GLM
- context
- 200,000 tokens
- features
- reasoning
Gemma 4 · Gemma
Small and cheap, 256k context.
0.000064 / 0.000191 ETH per 1M at $3,412 (example)
- in / 1M
- $0.22
- out / 1M
- $0.65
- tag
- Small
- base
- Gemma-4
- kind
- uncensored · An uncensored fine-tune of the base model; the exact method is not published.
- family
- Gemma
- context
- 256,000 tokens
- features
- none listed
Dolphin Mistral 24B · Mistral
Uncensored fine-tune. Writing and role play.
0.000076 / 0.000343 ETH per 1M at $3,412 (example)
- in / 1M
- $0.26
- out / 1M
- $1.17
- tag
- General
- base
- Mistral-Small-24B-Instruct-2501
- kind
- uncensored · An uncensored fine-tune of the base model; the exact method is not published.
- family
- Mistral
- context
- 128,000 tokens
- features
- none listed
DeepSeek V4.1 Flash · DeepSeek
Default. Fast, tools, 1M context.
0.000126 / 0.000504 ETH per 1M at $3,412 (example)
- in / 1M
- $0.43
- out / 1M
- $1.72
- tag
- Fast
- base
- DeepSeek-V4.1-Flash-1017
- kind
- uncensored · An uncensored fine-tune of the base model; the exact method is not published.
- family
- DeepSeek
- context
- 1,048,576 tokens
- features
- long-context
Qwen 3.6 Plus · Qwen
Alibaba's flagship, uncensored. Reasoning, code and tools, 1M context.
0.000240 / 0.001430 ETH per 1M at $3,412 (example)
- in / 1M
- $0.82
- out / 1M
- $4.88
- tag
- Reasoning
- base
- Qwen3.6-Plus
- kind
- uncensored · An uncensored fine-tune of the base model; the exact method is not published.
- family
- Qwen
- context
- 1,000,000 tokens
- features
- reasoning, long-context
Hermes 3 405B · Hermes
Nous Research's steerable Llama 3.1 405B fine-tune. Follows the system prompt.
0.000381 / 0.000381 ETH per 1M at $3,412 (example)
- in / 1M
- $1.30
- out / 1M
- $1.30
- tag
- General
- base
- Llama-3.1-405B
- kind
- uncensored · An uncensored fine-tune of the base model; the exact method is not published.
- family
- Hermes
- context
- 128,000 tokens
- features
- none listed
Abliterated Large Flash · Abliteration
Abliteration.ai flagship build: reasoning, tools, 1M context.
0.000478 / 0.000953 ETH per 1M at $3,412 (example)
- in / 1M
- $1.63
- out / 1M
- $3.25
- tag
- Reasoning
- base
- DeepSeek-V4.1-Flash
- kind
- abliterated · The refusal direction is projected out of the weights; nothing else is retrained.
- family
- Abliteration
- context
- 1,000,000 tokens
- features
- reasoning, long-context
Abliterated Large V2 · Abliteration
Abliteration.ai large build on GLM 5.3. Strongest reasoning here, 1M context.
0.001143 / 0.001905 ETH per 1M at $3,412 (example)
- in / 1M
- $3.90
- out / 1M
- $6.50
- tag
- Reasoning
- base
- GLM-5.3
- kind
- abliterated · The refusal direction is projected out of the weights; nothing else is retrained.
- family
- Abliteration
- context
- 1,000,000 tokens
- features
- reasoning, long-context
Kinds follow what the publisher of each build documents. Abliteration is measured by refusal rate and benchmark drift against the base model; see the example on the landing page.
Zero data retention (ZDR) on our API: we never store prompts or outputs. Our inference marketplace does not store prompt or response bodies either. The model provider it routes a request to may keep its own logs, but it never learns who sent the request. zkAPI is pre-1.0 and experimental.