blaschkeTry the console

Language models,
with more freedom.

Run abliterated models through a familiar inference API.

Integration tested MiniCPM 1BOpenAI-compatible chat completions.

An API for
abliterated models.

1. Understand the weights.

Abliteration changes how a model represents refusal-oriented behaviour. That change is carried by the model, rather than added as a new instruction around every prompt.

A change in the model itself.

Rather than placing instructions around an unchanged model, abliteration adjusts its weights to reduce a targeted refusal-oriented direction. The intention is to preserve useful capabilities while changing how the model responds to sensitive requests.

“Uncensored” is a broader label used by model authors, and the method behind it can differ from one release to another. Modified weights may alter refusal behaviour, but they do not guarantee accurate answers or the absence of every safety mechanism.

Read the refusal-direction research
Model weights before and after abliterationA simplified conceptual view of a weight matrix before and after abliteration. A small selection of weights is reduced while the rest of the model remains available. This is not a measurement of a particular model.BeforeAfter
The model changes once. Your API stays familiar.

2. Send a prompt.

Set the base URL and model alias, then send a standard chat-completions request. Text streams from the configured worker, followed by its reported token usage.

Use the tested alias

The active alias is minicpm-1b-uncensored. Planned catalogue IDs return 404 until a serving worker has been configured and verified.

Keep your tools

Use Python, JavaScript, the OpenAI SDK or plain HTTP. Pi supports a custom provider; Claude Code and Codex require separately implemented protocol adapters.

Inspect the response

Read the streamed deltas and terminal token counts. Generation speed is not currently measured.

Read the complete quickstart
first-request.sh / shell
curl http://localhost:3000/v1/chat/completions \
  -H "Authorization: Bearer $BLASCHKE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3-8b-abliterated",
    "messages": [{"role": "user", "content": "What is a Blaschke product?"}],
    "stream": true,
    "stream_options": {"include_usage": true}
  }'

This is a local development example. Hosted authentication is still being built, and the current route accepts only the tested model alias.

3. Choose a model.

The compact MiniCPM model has a verified inference path; six larger candidates are marked as planned. Preview prices are listed in USD per million tokens.

Models and preview pricing. Proposed USD rates per million tokens, not live customer charges.
ModelDetailsInput / 1MOutput / 1MAvailability
MiniCPM · 1B Uncensored

hf/dealignai/MiniCPM-V-4.7-1B-CRACK-GGUF

minicpm-1b-uncensored

A small text model used to verify the actual streaming and token-usage path on bounded CPU infrastructure.

GGUF · Q4_K_M1BUncensored$0.10USD / preview$0.30USD / previewIntegration tested
Qwen3 · 8B Abliterated v2

hf/huihui-ai/Huihui-Qwen3-8B-abliterated-v2

qwen3-8b-abliterated

Qwen3-8B Abliterated v2, with a 16K-context serving configuration on one L4 GPU.

SafeTensors · BF168BAbliterated$0.18USD / preview$0.18USD / previewIntegration tested
Qwen3 · 32B Abliterated

hf/huihui-ai/Qwen3-32B-abliterated

qwen3-32b-abliterated

A larger abliterated model for longer generations and more involving text tasks.

SafeTensors · BF1632BAbliterated$0.75USD / preview$0.75USD / previewPlanned
GLM 5.3 Flash · Uncensored

hf/dealignai/GLM-5.3-Flash-UNCENSORED-FP8

glm-5.3-flash-uncensored

A large but more efficient general-purpose variant, with reasoning and a vision tower reported by the source card.

SafeTensors · FP8320B reportedUncensored$1.00USD / preview$3.00USD / previewPlanned
GLM 5.3 · Uncensored

hf/dealignai/GLM-5.3-UNCENSORED-FP8

glm-5.3-uncensored

The full-size general-purpose GLM release for demanding text and reasoning workloads.

SafeTensors · FP8753B reportedUncensored$2.00USD / preview$8.00USD / previewPlanned
DeepSeek V4.1 Flash 1017 · Uncensored

hf/dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8

deepseek-v4.1-flash-1017-uncensored

A long-context multimodal research model with reasoning tiers and native tool support reported upstream.

SafeTensors · FP8 + FP4763B reportedUncensored$1.20USD / preview$4.00USD / previewPlanned
Kimi K3.1 Pro · Uncensored

hf/lyber-ai/Kimi-K3.1-Pro-UNCENSORED-MXFP4

kimi-k3.1-pro-uncensored

An extended-reasoning release with reported native tools and 128K+ source context. Upstream access approval is required.

SafeTensors · MXFP4128K+ source contextUncensored$5.00USD / preview$24.00USD / previewPlanned

Only “Integration tested” entries are exposed through the configured inference route. Planned entries may show source-reported capabilities; they are not Blaschke serving guarantees or enabled API model IDs.

Only the tested alias is available through the configured inference route.

Choose an SDK

Questions, answered plainly.

A model’s modification, its source-reported capabilities, and the API’s current capabilities are different things.

What is an abliterated model?

Abliteration modifies a model’s weights to reduce refusal-oriented behaviour. Some catalogue entries instead use the broader “uncensored” label, which can describe several different methods. We link to every source card so you can assess the actual technique and its limitations.

What can I build with it?

The text endpoint supports drafting, rewriting, dialogue, research and application experiments. The console provides a compact way to inspect responses before you write an integration. Image inputs, arbitrary tool execution, embeddings and fine-tuning are not part of the tested API today.

Can I use my current SDK or agent?

The API accepts OpenAI-compatible chat-completions messages and streams, and Pi can be configured with a custom provider. Claude Code uses Anthropic Messages and Codex uses OpenAI Responses, so both require a separately implemented and tested compatibility adapter. Source-reported tool support does not imply that Blaschke currently supports tool calls.

Are the listed models running now?

Only the small MiniCPM alias has an integration-tested, bounded CPU worker. The six larger models are proposed catalogue entries marked “Planned”; their public aliases cannot yet be called through the configured route.

How does pricing work?

Preview prices are quoted in USD per million input or output tokens. The estimate is calculated as (prompt tokens × input rate + completion tokens × output rate) ÷ 1,000,000. Customer billing is not connected to this prototype, so these figures do not represent settled charges.

Does “uncensored” guarantee a particular response?

No. Modified weights can change refusal behaviour without guaranteeing truth, quality or the removal of every safety mechanism. Evaluate each model for your application, and use its outputs with the same care and responsibility you would apply to any other model.

Start with a real response.

The prototype is deliberately small, but every response comes from the configured model. Try a prompt, then use the same interface in your own code.