Google Vertex Adds Mirendil's Self-Evolving RFT Architecture

Mirendil has secured a $100 million compute partnership with Google Cloud to scale its self-improving AI models for autonomous scientific discovery. The multi-year agreement migrates Mirendil’s primary training workloads to Google’s Hypercomputer architecture, granting access to TPU v6p and NVIDIA H200/B200 instances on A3 Ultra GPU clusters. The estimated $125 million investment over three years moves the startup’s focus from theoretical benchmarks to applied closed-loop discovery.

The Recursive Feedback Transformer

The compute resources are entirely dedicated to Mirendil-Alpha v2.1, a model family built on the Recursive Feedback Transformer (RFT) architecture. Traditional language models generate sequential tokens based on static weights, but RFT introduces a synthetic critic layer. This mechanism evaluates and regenerates the model’s internal weights in real time during the inference process, allowing it to autonomously correct logical errors.

To feed this architecture, Mirendil is scaling its internal Omniscience dataset from 15 trillion to 50 trillion tokens by Q1 2027. The training data focuses exclusively on hard scientific domains, including peer-reviewed literature, chemical synthesis logs, and protein folding data. The goal is to achieve autonomous hypothesis generation, simulation, and world model refinement without human intervention.

Vertex AI Integration and Benchmarks

Google Cloud has integrated Mirendil-Alpha into Vertex AI as a Private Preview for select pharmaceutical and materials science enterprises. The deployment leverages specialized Titanium offload engines to reduce networking overhead during massive parallel training and inference sessions.

The RFT architecture demonstrates measurable advantages in specialized domains. According to a technical paper published on August 8, Mirendil-Alpha v2.1 outperformed GPT-5 (Legacy) and Claude 4.5 by 14.2% on the Sci-Eval 2026 benchmark. The performance delta was particularly distinct in organic chemistry and quantum electrodynamics.

Primary training runs are currently split between Google Cloud’s us-central1 and europe-west4 regions. The partnership includes a carbon-neutral compute mandate, relying heavily on geothermal-powered infrastructure in Nevada to offset the high energy demands of recursive self-improvement algorithms.

Security and Alignment Constraints

The self-improving nature of the RFT architecture introduces distinct validation challenges. Because the model dynamically updates its own reasoning pathways during execution, static safety evaluations degrade rapidly. Researchers have raised concerns regarding the potential for reward hacking or the autonomous generation of biologically hazardous compound formulas.

To mitigate these risks, the current production build restricts output modalities and operates strictly within specialized scientific domains. The self-correction loops lack generalized access to arbitrary code execution, narrowing the surface area for unexpected autonomous behavior.

If you are evaluating AI agents for scientific or highly technical domains, the RFT architecture shifts the standard tradeoff between latency and accuracy. The real-time synthetic critic increases per-token computation time, but it significantly reduces the need to build external multi-agent coordination patterns for logic verification.