DEINORUS: 100% Subquadratic | Recall=Dense @1M | 2.80x

The description states that DEINORUS is an inference-time attention mechanism designed to reduce the quadratic attention barrier in large language models. It claims to do so without retraining the base model and without hiding a dense quadratic fallback inside its attention path.

It includes:

• A canonical PyTorch/Triton implementation
• A llama.cpp port with recall, reasoning, integration, and speed validation
• An experimental vLLM plugin validated for retrieval parity with dense through 128K
• Portability evidence from successful routing tests on Kimi Linear

The headline performance result is 2.80x faster than PyTorch SDPA at 512K tokens, measured on one NVIDIA H200.

This is a technical artifact, not a commercial product or service.

The description states that DEINORUS aims to solve the problem of quadratic attention growth with context length, which becomes expensive as models process more documents, code, history, or memory.

It positions itself as:

• Training-free
• Fully reproducible
• Dense-equivalent retrieval at scale (up to 1 million tokens)
• Subquadratic attention mechanism (not including unrelated operations in the surrounding model/runtime)

There is no indication of a shift from research to commercial positioning. The project remains framed as an engineering innovation, not a product with a go-to-market strategy.

Not evidenced.

The description does not identify any specific customer segments or ideal customer profiles (ICP). It focuses on technical performance rather than user needs or market targeting.

Not evidenced.

There is no mention of pricing, licensing, monetization strategies, or business model in the self-reported description. The project appears to be a research contribution with no commercial elements described.

The description states:

• DEINORUS includes working implementations in PyTorch/Triton, llama.cpp, and experimental vLLM
• It supports sparse prefill and decode operations
• It has been tested on Qwen and Kimi Linear architectures
• Performance gains are claimed at 2.80x faster than stock PyTorch SDPA at 512K tokens
• The work was enhanced using GPT-5.6 and Codex during OpenAI Build Week

These are technical claims, not signals of product maturity or commercial delivery.

Not evidenced.

There is no evidence of:

• Customers
• Revenue
• Adoption
• Product-market fit
• Market traction beyond the author’s own testing and validation

The project is described as a research prototype with experimental integrations, not a deployed product or service.

Not evidenced.

The description does not reference existing tools, platforms, or competitors in the attention mechanism or long-context LLM space. It does not compare DEINORUS to other solutions or describe its competitive positioning.

• No commercial traction: The project is self-reported and lacks any evidence of real-world use or adoption.
• Unverified claims: Performance results are presented as validated but without independent corroboration or third-party audits.
• Research-only focus: No indication of productization, scalability beyond experimental setups, or roadmap toward production deployment.
• Single-founder project: The team size is listed as one member, suggesting limited resources for scaling or commercial execution.
• Limited portability evidence: While tests were run on two model families (Qwen and Kimi Linear), this does not prove broad compatibility or generalizability.

1. What specific use cases or applications are you targeting with DEINORUS?
2. How do you plan to validate performance across a wider range of hardware, model architectures, and real-world workloads?
3. Are there any known limitations or edge cases in the current implementation that could affect scalability or reliability?
4. What is your roadmap for moving from experimental research to production-ready deployment?
5. Have you considered how DEINORUS might integrate into existing LLM serving infrastructure (e.g., Hugging Face, vLLM, Llamafile)?
6. How do you intend to monetize or commercialize this technology?