Learn
Plain explanations of the ideas behind MagmaRouter: what abliteration is, how a spend cap can actually hold, what zero prompt retention does and does not promise, and what makes an API OpenAI- or Anthropic-compatible.
AbliterationAbliteration removes the one direction in a model’s activations that produces refusals, by editing the weights, with no retraining. How it works, step by step, and what it costs.
Refusal directionA single direction in a model’s residual stream carries its tendency to refuse. Finding it is the basis of abliteration.
Uncensored vs abliteratedAbliterated models had the refusal direction removed, so they rarely refuse. Uncensored ones were finetuned to refuse less, and still sometimes do. Side by side.
Why LLMs refuseRefusals are trained in after pretraining, by instruction tuning and preference training, and some are added by the platform in front of the model. Where each one comes from and how to tell them apart.
Ways to uncensor an LLMAbliteration, uncensored finetunes, LoRA adapters, model merges, base models and jailbreak prompts all reduce refusals. What each one changes, what it costs, and how long it lasts.
Abliteration: pros, limits, risksAbliteration cuts false refusals cheaply and keeps the original instruction tuning, but it can cost capability, leave some refusals in place and make a model too agreeable. What to expect before you rely on one.
Heretic: automatic abliterationWhat Heretic is, how it finds abliteration settings by itself, what its own results show, why models named heretic keep appearing on Hugging Face, and an honest answer to whether it is as scary as it sounds.
Evaluating abliterated modelsTest an abliterated model for what it gained and what it lost: refusal rate on held-out prompts, indirect refusals, capability against the original, agreeableness and tool calling. With a script.
Running abliterated models locallyRun an abliterated GGUF on your own machine with Ollama or LM Studio, pick a quantization that fits your memory, and know when a hosted API is the better trade.
Well known abliterated modelsThe community abliterated releases people search for most: Llama 3.1 8B, Qwen3 14B, Gemma 3 27B, Mistral Small 3.2 and DeepSeek R1 Distill 32B. Who made each, what it is based on, and what it keeps.
Uncensored LLM APIAn uncensored LLM API serves models that do not refuse, with no moderation layer in front. Where refusals actually come from, how to test an API before you trust it, and what to check.
API without a content filterSome LLM APIs run a classifier on every prompt and reply and block what it flags. Which ones do by default, how that differs from a model refusing, and what an unfiltered API really means.
OpenAI base URLThe OpenAI base URL is https://api.openai.com/v1 by default. How to point the Python and Node SDKs, environment variables, LangChain, LlamaIndex, the Vercel AI SDK and LiteLLM at another endpoint.
LLM gatewayAn LLM gateway is one API in front of many model providers, adding keys, billing, spend limits and logging without changing your code.
Per-key spend capsA spend cap that is checked before the call, not after, so a runaway or leaked key stops at the number you set instead of billing past it.
max_tokens clampingMagmaRouter lowers the outbound max_tokens to whatever the remaining budget covers, so a response cannot cost more than was reserved.
Zero data retentionMagmaRouter keeps zero data retention: it stores the model, token counts and cost, never the prompt or the answer, encrypts every hop and returns a signed receipt so you can verify it.
How your messages are protectedEvery request is encrypted in transit, never written down, and answered with a signed receipt that holds only a fingerprint of what you sent. Zero data retention, step by step.
RefusalsWhat a model refusal is, the two ways the Xploded line has it taken out of the weights, the work other APIs refuse, and what an Xploded model can still say no to.
OpenAI-compatible APIAn OpenAI-compatible API accepts the same requests as OpenAI’s chat completions endpoint, so any OpenAI SDK works by changing the base URL. Working curl, Python and Node examples.
Anthropic-compatible APIAn Anthropic-compatible API accepts the Messages format, so Claude Code and the Anthropic SDKs work by changing the base URL.
Signed receiptsMagmaRouter returns an HMAC-signed receipt of the prompt hash on every call, so you can prove later what the gateway bound to without it keeping the prompt.
Red-teaming LLMsRed-teaming an LLM means attacking your own model on purpose: jailbreaks, prompt injection and abuse generation, to find failures before an attacker does.
DeepSeek V4.1 Flash 1017DeepSeek V4.1 Flash 1017 is the 17 October 2026 update to V4.1 Flash. What DeepSeek changed, its own benchmark table against the September release, and what it means if you want it uncensored.
DeepSeek V4.1 Flash vs V4 FlashDeepSeek V4.1 Flash replaced V4 Flash in September 2026: a new architecture, native vision and higher agent scores. The timeline, DeepSeek's own numbers and the Artificial Analysis scores side by side.
DeepSeek V4 Pro vs V4.1 FlashDeepSeek's Flash model now scores above its Pro model. DeepSeek V4 Pro against V4.1 Flash on Artificial Analysis and on DeepSeek's own published benchmarks, and where V4 Pro still leads.
DeepSeek V4.1 Flash APICall DeepSeek V4.1 Flash 1017 Xploded, the uncensored DeepSeek V4.1 Flash, with curl, Python or JavaScript. Base URL, model id, a working tool call and the limits that matter.
DeepSeek V4.1 Flash reasoning effortHow reasoning effort works on DeepSeek V4.1 Flash: the 1 to 100 scale in the weights, low, high and max on DeepSeek's API, what Artificial Analysis measured with reasoning on and off, and how to set it.
Codex CLI on DeepSeek V4.1 FlashRun OpenAI Codex CLI on DeepSeek V4.1 Flash 1017 Xploded through the MagmaRouter Responses endpoint: the config.toml provider block, what is translated and what is not.
DeepSeek V4.1 Flash context windowDeepSeek V4.1 Flash has a 1,048,576-token context window, and the uncensored build on MagmaRouter writes up to 262,144 tokens in one reply. How the two limits interact and how they compare across the line.
Kimi K3.1 ProKimi K3.1 Pro scores 55 on the Artificial Analysis Intelligence Index, fourth of every model at launch. What Moonshot AI announced, where the weights stand, and what to use uncensored today.
Kimi K3.1 Pro APIIs there a Kimi K3.1 Pro API? Not at launch: Moonshot AI announced the model without a public API or weights. What that means for agents and apps, and an uncensored model to call through the same OpenAI SDK today.