AENEA Global has open-sourced BHR-0.1, a 74.8M parameter System 1 decision engine built on the novel CartanLM architecture with Block Householder Linear layers constraining weights to the Stiefel manifold. Evaluated on a 250-question specialist benchmark against Laya (421M), BHR-0.1 achieves 98.0% contradiction detection (vs. Laya's 82.2%), 100% accuracy on procedural JSON rule execution, and an elite 0.0017 Brier score, all while running at 187.9ms average latency on CPU—1.7x faster than the 421M baseline. The model was trained via KL Divergence against soft target distributions on a rigorously audited 538,238-record dataset spanning Choice, YesNo, and Score decision primitives. BHR-0.1 ships with fp32 PyTorch weights, a custom QT-Cittern-12k tokenizer, a production-grade FastAPI inference engine, and a fully containerized Docker image for zero-dependency edge deployment. The orthogonal geometry preserves semantic boundaries that dense models collapse, proving that orthogonal attention scales down better than dense transformers for enterprise routing and triage. Released under Apache 2.0.
AENEA Global has open-sourced Pinta-2.0, a 74M parameter semantic router built on the 24-layer Cittern-2.0 architecture with block Householder orthogonal attention and a custom 12,000-token QT-Cittern-12k vocabulary. Evaluated on the 1,060-prompt Balanced Routing Benchmark v6 (zero train overlap, English-only), Pinta-2.0 achieves 92.45% strict accuracy and 0.92 macro F1 across nine reserved domain tokens, with near-perfect Compliance routing (0.99 precision) and flawless Math recall (1.00). The model was fine-tuned from the Cittern-2.0 base checkpoint over three epochs on a rigorously audited 57,968-record dataset using an effective batch size of 64, a learning rate of 1e-5, and orthogonal regularisation to preserve the Stiefel manifold geometry. Pinta-2.0 ships with fp32 and fp16 PyTorch weights, an ONNX fp32 graph for cross-platform inference, a zero-dependency C++ edge router, and a Python ONNX reference implementation. Trained on consumer hardware (NVIDIA RTX 4090). Released under Apache 2.0.
AENEA Global has open-sourced Pinta-1.1, a 226M parameter high-speed semantic router built on the custom Merlin v2 architecture and trained on top of the AENEA Prelude-6 base checkpoint. Designed as an ultra-lightweight MoE gateway (approximately 656 MB in bfloat16), Pinta-1.1 categorises incoming prompts across a high-precision domain taxonomy via nine reserved special tokens spanning Code, Mathematics, Document QA, Architecture, Creative, and Knowledge. Evaluated on the 1,020-prompt Pinta Gold benchmark, the model achieves 84.61% strict accuracy and 87.65% flexible accuracy with an average routing latency of approximately 192 ms. The release includes the Pinta Router Engine, a production-ready Python SDK and FastAPI middleware providing an OpenAI-compatible completions endpoint with configurable confidence fallback logic, streaming proxy support, and native async dispatch to Ollama, vLLM, the OpenAI API, or local Hugging Face pipelines. The Merlin v2 architecture features 16 transformer blocks with parallel residuals, Grouped-Query Attention (16Q/4KV), SwiGLU FFNs, and RoPE positional embeddings paired with the 49,152-vocabulary QT VI.6.4 UltraLingo tokenizer. Pre-trained on a high-density technical corpus and fine-tuned on a 50/50 mix of curated imperative data with hard-negative synthetic boundary pairs distilled via DeepSeek-v4-Flash, entirely on consumer hardware (RTX 4090 / RTX 4060). Released under Apache 2.0.
AENEA Global has released the QT Generation VII Tokenizer Family — Ensemble, Overture, and Prelude — now available on Hugging Face under the Apache 2.0 license. The release supersedes the QT.VI generation with a tiered suite of byte-level BPE tokenizers covering 204 FLORES-200 languages, programming syntax, and scientific notation. Ensemble (QT.VII.1.8.Plus, 129K vocabulary) targets frontier models from 8B to 70B+ parameters with universal language support and zero script collapse, leading the family in human-language compression at 2.228 characters per token and statistically tying DeepSeek V4 on Python code. Overture (QT.VII.1.8.Mini, 96K) serves as the mid-tier build for 1B–8B parameter models, cutting embedding VRAM by 25.7% versus 129K vocabularies while keeping code sequence expansion under 1.5%. Prelude (QT.VII.1.8.Micro, 64K) targets micro-models at or below 1B parameters, improving pure code density by over 19% compared to the legacy 64K QT.VI.3 build. Benchmarked across TokenizerBench, a Python code corpus, and FLORES-200 regional script health checks, the QT builds avoid the byte-fallback spikes that commercial baselines exhibit on regional scripts such as Amharic, Georgian, Armenian, Kannada, and Lao.
Quartz has open-sourced QT VI.3.5 UltraLingo, a 64K tokenizer that becomes the new flagship of the QT line, built for 500M-1B parameter models. It covers all 204 FLORES-200 languages losslessly where Llama 2 and Mistral fall back to raw bytes on 34, and beats both on 11 of 13 major languages (tying Llama 2 on German) at a global mean of 3.35 bytes per token - roughly 41% more compression than either. VI.3.5 improves on the 32K VI.1.3 across 198 of 204 languages, roughly doubling or tripling the industry 32K tokenizers on CJK, Arabic, and Hindi. English is the one language it does not win: a 64K tokenizer covering 204 languages cannot out-compress a 32K tokenizer that spends its whole vocabulary on English, so VI.3.5 reaches parity with the 32K flagship (3.84 vs 3.90 characters per token) but trails Llama 2's 4.24 by design - the honest, structural cost of universal coverage. VI.1.3 32K remains the flagship for smaller models. Released under Apache 2.0.
Quartz has open-sourced QT VI.1.3 UltraLingo, the new Prelude-tier flagship tokenizer for the AENEA model series. At the same 32K vocabulary budget as Llama 2 and Mistral, VI.1.3 covers all 204 FLORES-200 languages losslessly across 27+ writing systems with zero coverage failures - where both industry tokenizers fall back to raw bytes on 34 languages each. It beats Llama 2 on 107 of the 170 languages Llama can tokenize at all, and Mistral on 137, roughly doubling their efficiency on Arabic, Hindi, and Thai while leading on CJK. The build fixes the English over-fragmentation of its predecessor, QT V.4.6, reaching 1.53 tokens per word and 3.90 characters per token on English, and is more equitable overall with a 6.4x cross-lingual spread versus 7.6x. The documented cost is a modest ~8% English trade against Llama 2, the honest price of universal coverage at a fixed vocabulary; the planned 64K Overture tier targets the low-resource complex scripts without that trade. Released under Apache 2.0.
The Crassus research platform has reached its 10th and 11th published reports with two companion pieces: "The Sports Sponsorship Signal," identifying a non-financial behavioural leading indicator of credit sector stress with a 100% historical hit rate across the 2004-2008 financial crisis, and "The Consumer Credit Ecosystem," mapping the $200 billion pipeline from fintech origination through private credit absorption to pension fund destination. Both reports utilise the Crassus provenance badge system (Confirmed, Reported, Derived, Modelled, Signal) and have been independently fact-checked against public filings, press releases, and regulatory data. The research library now spans structured finance, consumer credit, insurance, infrastructure, and original behavioural analysis.
The Crassus division has launched the Crassus National Intelligence Terminal, a real-time intelligence platform covering 250 countries, 14 commodity markets, 30 economic indicators, and 30+ international news sources. Built as a mobile-first progressive web app on free public APIs (World Bank, IMF, FRED, BBC, Al Jazeera, The Guardian, Reuters, and others), the platform is engineered for governments, sovereign wealth funds, and institutional clients with no per-seat licensing. The Crassus brand has been formally restructured into two terminals: the new National Intelligence Terminal and the existing V5.0 Financial Terminal, sharing a common engineering foundation.
AENEA Prelude-5 Reaches Deep Factual Crystallisation Zone
Prelude-5 (276M parameters, trained on the new QT V.4.4 32K UltraLingo tokenizer) has passed step 40,000 with loss 2.249, perplexity 9.48, and gradient norm 0.208 - reaching the deep factual crystallisation zone five times faster than prior generations. This validates the Factual Crystallisation Hypothesis: gradient norm rather than loss is the primary predictor of factual recall emergence in language models trained on ultra-clean data.
Quartz has open-sourced QT V.4.4 32K UltraLingo, the new flagship tokenizer for the Prelude model series. Covering 72 languages across 27 writing systems with a 31.5x cross-lingual equity ratio - the highest of any QT generation - the V.4.4 achieves 4.231 fertility and consumes 15.7% fewer total tokens than Llama 3 across the FLORES-200 benchmark. The companion QT V.4.1 64K UltraLingo, released in parallel, serves the larger Overture model family targeting 500M-2B parameter scales. Both are published under Apache 2.0.
The Crassus division has expanded its published research portfolio to eight reports covering Thames Water liquidity restructuring, consumer ABS origination-to-default chains, BNPL systemic risk, European SRT markets, US CRT/SRT analysis, insurance and reinsurance systemic risk, de-SPAC convertible debt maturity walls, and hidden vulnerabilities in consumer credit systems. All reports are available at crassus.info.
The Crassus platform has deployed a comprehensive options analytics suite featuring 3D implied volatility surfaces, put-call ratio analysis, gamma exposure with flip point detection, earnings move analysis, and unusual activity detection across US equity markets. The terminal comprises 8 backend endpoints, 23 Pydantic models, and 19 frontend components.
Prelude-4 (276M parameters, d=1024, 16 layers, GQA 4:1) has exceeded the factual recall performance of Prelude-1, reaching its target loss and gradient norm zone significantly faster. This training run has contributed to the development of the Factual Crystallisation Hypothesis - a novel framework proposing that gradient norm, not loss, predicts the onset of factual recall capabilities in language models.
Quartz and AENEA have open-sourced the new QT V.3 32K UltraLingo, a parameter-efficient SuperBPE multilingual tokenizer designed for sub-500M parameter models. Covering 71 languages across 26 writing systems within a highly compact 32,000-token vocabulary, it utilizes innovative two-stage SuperBPE training. The V.3 UltraLingo delivers 3× better equity than industry standards and outperforms models 3-4× its size on Indic and SE Asian scripts, freeing up between 65M and 262M embedding parameters to be redeployed into reasoning layers.
Quartz and AENEA have jointly open-sourced the second-generation tokenizer family, featuring 64K, 96K, and 114K Code variants. Validated across 204 languages on the rigorous FLORES-200 benchmark, the QT_V.2 suite achieves the fewest total tokens in its class while demonstrating a 4× improvement in cross-lingual equity compared to industry standards.
AENEA has successfully trained and released Prelude-1, a highly-optimised 284M parameter base model. Trained exclusively on Quartz-cleaned corpora, Prelude-1 proves the company's foundational hypothesis: smaller models trained on surgically clean data can consistently outperform networks up to 3× their size in factual recall and precision.
The Quartz data division has published its SE Ultra-Clean pipeline. Transforming 23 distinct Stack Exchange communities into high-quality, instruct-format Q&A pairs, this pipeline expands the AENEA training substrate into complex procedural reasoning and technical problem-solving domains.
Quartz Achieves Sub-8-Hour Wikipedia Processing with Pipeline v6
In a major infrastructure milestone, Quartz Pipeline v6 successfully processed the entire English Wikipedia dump in under 8 hours. Utilising parallel decompression, regex XML splitting, and pre-computed MinHash signatures, this engineering feat establishes a new standard for rapid, deterministic data curation at scale.
AENEA Global Ltd announces the formal structuring of Crassus, a dedicated quantitative research division. Crassus will focus exclusively on the systematic analysis of distressed credit, insured infrastructure, and structured credit cascades, leveraging the factual baseline models developed by AENEA.
AENEA Global Ltd has been formally incorporated and registered in Manchester, United Kingdom. The holding company is established to build vertically integrated AI infrastructure, uniting data curation, model training, and applied quantitative research under a single architectural philosophy.