Abliterated Large Flash

• 🏢 Abliteration AI's unrestricted text-and-vision model, released October 2026.
• 🧠 Based on DeepSeek V4.1 Flash, then further abliterated.
• 📏 One-million-token context window, served at FP8 quantization.
• 🔧 552B backbone; 8B active on input, 16B on output.
• 👁️ Natively processes images alongside text input.
• 🎯 Function calling, web search and code-oriented use listed.
• 🔒 Vendor states no content-filtering guardrails.
• 🆕 Newest release in Abliteration's catalog, after Abliterated Large V2.

Abliteration AI

Abliteration AI is a Palo Alto company selling an OpenAI-compatible LLM API for work most assistants turn down: AI red teaming, cybersecurity, trust and safety, synthetic data and evals, ML research, and defense or government workflows. Its models start from…

Abliterated Large Flash is Abliteration AI's multimodal, unrestricted model, listed with a release date of 23 October 2026. According to the provider's model documentation at docs.abliteration.ai and the catalog listing, it is based on DeepSeek V4.1 Flash and further abliterated, meaning the base model's refusal behaviour is removed at the weight level so the deployed model ships without content-filtering guardrails.

On the architecture and serving side, the provider's documentation describes a 552-billion-parameter backbone operating sparsely, with roughly 8 billion parameters active on input and 16 billion on output, and a context window of one million tokens. The catalog lists FP8 as the serving quantization on this hosting platform. Declared capabilities include reasoning, native image-and-text input, code-oriented generation, function calling and web search.

Within Abliteration AI's own line-up, it follows Abliterated Large V2 from earlier in October 2026. The two sit in separate family lines rather than being successive versions of the same model: Flash is the DeepSeek-V4.1-Flash-derived branch, and it adds native vision input plus the one-million-token window in a sparsely activated configuration.

No license is listed for this model in the catalog. Unaffiliated community abliterations of DeepSeek V4.1 Flash, such as those published on Hugging Face, circulate separately and are not the same artifact.