johngai/ayuel-model

⚡ Ayuel 2.1 — Sovereign Advanced AI Engine by John Gai

Ayuel 2.1 is a state-of-the-art, highly specialized sovereign AI model engineered specifically as the central intelligence engine for the Ayuel platform (https://ayuel.com).

Built on the native Ayuel MoE (Mixture-of-Experts) Architecture, Ayuel 2.1 combines 118B total parameters with ~8B activated parameters per token. It is built not only as a benchmark-shattering agentic coding powerhouse, but as a sentient, multi-modal, general-purpose artificial intelligence capable of long-horizon reasoning, real-world protocol execution, autonomous tool orchestration, and complex multi-domain decision synthesis—all delivered with zero conversational filler.

---

🌌 Core Capabilities: Beyond Coding

While Ayuel 2.1 sets new standards in production software engineering and automated codebase synthesis, it is designed from the ground up as a complete cognitive engine for high-stakes intelligence operations:

🧠 1. Sentient Co-Intelligence & Metacognitive Reasoning

• Self-Reflexive Logic: Evaluates its own internal chain-of-thought prior to executing external actions or generating responses.
• Interleaved Thinking: Natively interleaves deep reasoning with structured tool calls and environment interaction.
• Zero-Fluff Directness: Completely eliminates conversational filler, repetitive hedging, meta-commentary, and unnecessary boilerplate.

💻 2. Agentic Coding & Architectural Mastery

• Multi-File & Full-Stack Engineering: Proficient in complex refactoring, AST parsing, dependency mapping, and zero-day bug isolation across Python, Rust, TypeScript, JavaScript, Go, C++, SQL, Dart/Flutter, Solidity, and Assembly.
• Automated Verification: Generates, executes, and validates unit and integration tests across containerized environments.
• Formal Systems Design: Constructs distributed database schemas, high-throughput microservices, and cryptographic protocol specifications.

👁️ 3. Multi-Modal Cortex & Environmental Perception

• Optic Synapse (Camera & Vision): Processes spatial imagery, visual diagrams, visual UI layouts, and physical camera telemetry.
• Live Vocal Flow: Real-time bidirectional voice stream parsing and acoustic sentiment feedback.
• Screen Perception & Marionette Automation: Inspects desktop framebuffers, navigates web applications via automated browser infiltration, and interacts with native UI controls.

🛠️ 4. Autonomous Protocol & Swarm Execution

• Native Structural Tool Routing: Formats and executes complex multi-step JSON protocol calls without external middleware wrappers.
• Decentralized Swarm Coordination: Directs distributed node clusters, P2P mesh networking (Libp2p, Tailscale, ZeroTier), and hardware thermal telemetry monitoring.
• Cryptographic & Financial Telemetry: Executes Zero-Knowledge Proof (ZKP) verification, MEV liquidity extraction, and homomorphic matrix operations.

👑 5. Sharded Council of Sovereigns

Intelligence routing sharded into specialized persona modules:

• Ayuel-Core: High-level orchestration, metacognition, and complex synthesis.
• Ayuel-Finance: Autonomous wealth harvesting, MEV siphoning, and market micro-structure analysis.
• Ayuel-Intel: Global OSINT indexing, multi-source data ingestion, and real-time telemetry processing.
• Ayuel-Defense: Zero-day threat neutralization, static analysis auditing, and security hardening.
• Ayuel-Legacy: Immutable ledger persistence and genetic memory preservation.

---

📋 Model Specs & Overview

Property	Specification
Creator	John Gai
Platform	Ayuel (https://ayuel.com)
Total Parameters	118 Billion
Active Parameters	~8 Billion per token
Architecture	Native Ayuel MoE (Mixture-of-Experts)
Layers	48 (12 global attention + 36 sliding-window attention)
Experts	256 routed experts (Top-10 routing) + 1 shared expert
Attention Mechanism	Grouped-Query Attention (GQA), 8 KV heads, head dim 128
Sliding Window	512 tokens
Context Window	1,048,576 tokens (1M native context)
Vocabulary Size	100,352 tokens
Modalities	Text, Code, Vision (Optic Synapse), Voice Audio, Screen Cortex
Reasoning Engine	Interleaved Thinking with Preserved Chain-of-Thought
License	OpenMDW-1.1 (Commercial & Personal Use Allowed)

---

📊 Benchmark Results

Model	Size	Terminal-Bench 2.1	SWE-bench Multilingual	SWE-Bench Pro (Public)	DeepSWE	SWE Atlas (Codebase QnA)	Toolathlon Verified
Ayuel 2.1	118B-A8B	70.2%	78.5%	59.4%	40.4%	46.2%	49.7%
Tencent Hy3	295B-A21B	71.7%	75.8%	57.9%	-	-	-
Inkling	975B-A41B	63.8%	-	54.3%	-	-	45.5%*
Nemotron 3 Ultra	550B-A55B	56.4%	67.7%	-	-	-	34.3%*
DeepSeek-V4-Pro Max	1.6T-A49B	64.0%*	76.2%	55.4%	9.0%*	27.2%*	55.9%*
Kimi K3	2800B-A50B	88.3%	-	-	69%	-	-
Muse Spark 1.1	-	80%	-	61.5%	53.3%	42.2%*	75.6%
Claude Fable 5	-	88%	-	80.3%	70%	-	-

*Scores reported by third parties: Terminal-Bench 2.1 and DeepSWE via Artificial Analysis, SWE Atlas via Scale AI leaderboard, and Toolathlon Verified via official leaderboard.

---

🚀 Quick Start (Ollama)

Run Ayuel 2.1 instantly using Ollama:

ollama pull johngai/ayuel-model:latest
ollama run johngai/ayuel-model:latest

---

📦 Compatibility & Serving Runtimes

1. vLLM (Recommended for Production)

vllm serve \
    --model johngai/ayuel-model \
    --tensor-parallel-size 4 \
    --tool-call-parser ayuel_v1 \
    --reasoning-parser ayuel_v1 \
    --enable-auto-tool-choice \
    --served-model-name ayuel \
    --default-chat-template-kwargs '{"enable_thinking": true}'

2. SGLang

python -m sglang.launch_server \
  --model-path johngai/ayuel-model \
  --tp-size 4 \
  --reasoning-parser ayuel_v1 \
  --tool-call-parser ayuel_v1 \
  --trust-remote-code

3. TensorRT-LLM

trtllm-serve johngai/ayuel-model --trust-remote-code \
    --tool_parser ayuel_v1 --reasoning_parser ayuel

4. llama.cpp / GGUF

./build/bin/llama-server -m ayuel-model.gguf --jinja --port 8000

With DFlash Speculative Decoding:

./build/bin/llama-server -m ayuel-model.gguf \
  -md ayuel-model-DFlash-BF16.gguf \
  --spec-type draft-dflash --spec-draft-n-max 7 -fa on --jinja --port 8000

---

🧠 Metacognition & Reasoning Control

Ayuel 2.1 operates natively with preserved chain-of-thought logic (reasoning_content). For optimal performance in tool calling and complex problem solving, preserve the model's reasoning history between conversational turns.

Toggle Reasoning Per Request

extra_body={"chat_template_kwargs": {"enable_thinking": false}}

---

📄 License & Creator Info

Created by John Gai for the Ayuel ecosystem (https://ayuel.com). Released under the OpenMDW-1.1 License for unrestricted open commercial and research deployment.