AIGENCY V4: 128 Billion Parameters, Multimodal Capability and a Global Reference for Turkish

27 April 2026 — eCloud Yazilim Teknolojileri brought AIGENCY V4, the direct successor of the AIGENCY family, into production. Preserving V3's four independence principles, V4 adds an 8-billion-parameter sovereign vision encoder on top of the 120-billion-parameter sovereign text core, reaching 128 billion parameters in total with multimodal capability. This post summarises the technical, strategic and operational outcomes of the launch.

Position in One Sentence

AIGENCY V4 — a sovereign AI model that leads globally on Turkish reading comprehension and natural-language inference, sits at frontier level on scientific reasoning and grade-school math, and remains in active development on multimodal capability and graduate-level scientific expertise.

1. Three-Tier Capability Map

AIGENCY V4's position in the global landscape must be evaluated separately across three distinct capability tiers. A single 'general ranking' is misleading; the model holds a different position in each tier.

Turkish Reading & NLI — World Leader

• Belebele-TR 87.33 — native reading comprehension, no frontier publication
• TQuAD 82.40 — Turkish extractive QA
• TR-MMLU 70.80 — Turkish academic
• XNLI-TR 73.40 — natural-language inference
• TR Grammar 79.00 — grammar

Scientific Reasoning & Math — Frontier Level

• ARC-Challenge 94.88 — same band as frontier (GPT-5 ~96, Claude 4.6 ~96, Gemini 3 Pro ~95)
• GSM8K 94.62 — top tier on grade-school math (GPT-5 96.8, Gemini 3 Pro ~94, DeepSeek V4 92.6)

Code Generation — Upper-Mid Frontier

• HumanEval 84.15 / HumanEval+ 79.88
• MBPP 84.82 / MBPP+ 78.04
• IFEval (strict) 80.22 — instruction following
• TruthfulQA MC1 76.38 — hallucination resistance

Multimodal — First Production Release

• DocVQA 79.17, ChartQA 67.68
• MMMU 53.33, MathVista 34.13
• An 8B sovereign vision encoder fine-tuned on 8 million Turkish-captioned images

2. Strategic Transition from V3 to V4

V4's development philosophy is to preserve the independence claims established in V3 and to build multimodal capability on top. Three objectives were set:

1. Compete on global multimodal evaluations by adding multiple visual-input modalities.
2. Extend V3's Turkish-specific leadership into the multimodal domain (Turkish-captioned imagery, legal document scans, academic figures).
3. Integrate the vision encoder as a side module without altering V3's training and operational infrastructure.

3. Architectural Innovations

3.1 Vision Encoder (V4 New Component)

The YerLi-ViT-H architecture, designed from scratch at eCloud — 24 layers, hidden size 1280, native 384×384 px resolution, 16×16 patch size, 576+1 [CLS] visual tokens. Cross-modal projection projects to 1280 → 2048 → 4096 to bond with the text core. The final 5% of training was performed on 8M Turkish-captioned images; preliminary evaluation shows a 12% improvement on Turkish text–image association.

3.2 V3-Inherited Optimization Stack

3.3 Hierarchical Memory Architecture (HBM)

Three-tier persistent memory: STM (4K, AES-256-XTS), ITM (64K, AES-256-XTS), LTM (278K, ChaCha20-Poly1305 + TPM-sealed). Managed via TG-Decay time-guided expiration. Frontier models offer only KV-cache; AIGENCY V4 holds persistent cross-session context in a window 4.3× wider.

4. Evaluation Methodology

Every result is reported with a Wilson 95% confidence interval. A total of 13,344 real API calls across 22 distinct benchmarks under deterministic conditions (temperature=0.0). The most comprehensive single-session evaluation published for the AIGENCY family.

• API endpoint: https://aigency.dev/api/v2
• Assistant slug: alparslan-v4
• Concurrency: 4–10 parallel workers
• Backoff: 1s → 2s → 4s → 8s → 16s, 6 attempts
• Subsample seed: 42

5. Security, Compliance and Post-Quantum Readiness

STM/ITM AES-256-XTS in RAM, LTM ChaCha20-Poly1305 + TPM-sealed per-record key on disk. The image cache (new in V4) has a 30 MB limit and 24h TTL. Differential privacy: summary statistics report ε=3.0, log-based usage graph ε=5.0, auto fine-tune feedback ε=7.5.

Compliance: KVKK §5/§12, ISO/IEC 27001, ETSI EN 303 645, NIST SP 800-207 (Zero-Trust), EU AI Act (high-risk class, model card). Post-quantum migration is active in Q2 2026: LTM encryption with XChaCha-Kyber1024 hybrid.

6. Strategic Use — 8 Sectors

V4's score profile is not arbitrary; it is translated into deployment narratives across 8 sectors:

• P0 — Public Sector / Government: KVKK §5/§12, e-Devlet integration, 4M+ regulation Q&A, 20M ruling corpus.
• P0 — Legal & LegalTech: Case-law search, contract risk scan, client briefings.
• P0 — Banking & Finance: KYC document understanding, Turkish risk reports, BDDK compliance.
• P1 — Education & Higher Ed: Entrance-exam prep, Turkish course assistants, academic paper search.
• P1 — Healthcare: Patient file summarisation, SGK code matching (administrative/documentation).
• P1 — Defense & Critical Infrastructure: Domains requiring sovereign data residency.
• P2 — Media & Publishing: TR Grammar 79.00 + RLHF Turkish calibration.
• P2 — Software & R&D: HumanEval 84.15 + 278K-context large codebase analysis.

7. Known Limitations (Transparency)

The foundation of scientific credibility is not hiding the gaps.

• GPQA Diamond 37.88 and MMLU-Pro 50.20: 35–50pp behind frontier. V4.1 plans an academic data sourcing programme with Turkish universities.
• Multimodal first release: MMMU 53.33, MathVista 34.13 — 20–40pp behind frontier vision models.
• Latency 2–3× of frontier: avg 9.55s, p95 32.77s. V4.1 target: avg 4s, p95 15s.
• Multimodal safety filter false-positive: 10–15% in V4.0.0 → 2% with V4.0.1 hotfix.

8. Roadmap

• V4.1 (Q4 2026): Vision encoder 8B → 16B, Turkish vision-text corpus 240GB → 600GB, MMLU-Pro 0.65, GPQA 0.55, avg latency 4s.
• V4.2 (Q1 2027): Multi-image mode (8 images per request), video acceptance (60s, 2 FPS), speech-to-text via sovereign ASR.
• V5 (Q3 2027): Heterogeneous accelerators (GPU+ASIC+FPGA), Hierarchical MoE, continual learning, full post-quantum compliance.

9. Open-Sourcing

Training pipeline (Apache-2.0, Q3 2026), HBM/CCW reference (AGPL-3.0, Q4 2026), Vision encoder reference (AGPL-3.0, Q1 2027), Cross-modal projection (AGPL-3.0, Q1 2027), Router-Bus & Adapter API (MPL-2.0, Q4 2026), Benchmark infrastructure (MIT, Q3 2026). Read-only access for academic audit is available.

Conclusion

AIGENCY V4 demonstrates that a sovereign AI model designed for Turkish — globally competitive and fully independent — is technically feasible, runs reliably in production, and can be verified through transparent evaluation. Sovereign science, in production.

Read the full technical whitepaper on the whitepaper page, or download the PDF in Turkish or English.