openmoe.ai: Diamonds, Breakthroughs & Open-Source Arsenal

A fact-checked intelligence report on the open-source AI governance stack. Verified 12 Jun 2026 by Mavis against live PyPI, GitHub, HuggingFace, MCP Registry, and meok.ai sources.

FACT-CHECKED   View source on GitHub   Download markdown   First edition (for diff)

1. The CSOAI / MEOK / Hives / OPENMOE Stack

1.1 CSOAI.org — The Governance Nerve Center VERIFIED

The Council for the Safety of AI (CSOAI) is a runtime-enforceable governance fabric. 478 public repositories on GitHub under CSOAI-ORG (verified via GET /users/CSOAI-ORG on 12 Jun 2026). 30 servers are listed in the official registry.modelcontextprotocol.io index when filtered for CSOAI (verified via the live registry API). The previous edition's "294 servers verified in the official MCP Registry" number is NOT what the live registry returns — 294 is an internal-org count, not a registry count.

Correction from previous edition: the crosswalk supports 13 frameworks (EU AI Act, ISO 42001, NIST AI RMF, GDPR, SOC 2, HIPAA, FedRAMP, CCPA, PIPEDA, DPDPA, LGPD, CSA, OWASP), not 30. The "30 frameworks" claim was unsupported; the actual MCP server README and PyPI summary both say 13.

EU AI Act timeline note: the meok-watermark-attest-mcp PyPI summary says the nearest deadline is 2 November 2026 (the Digital Omnibus pushed high-risk obligations to 2 December 2027). The previous edition's "Article 50 hits 2 August 2026" is stale; the live package documentation references the November 2026 cliff instead. Worth flagging with the legal team before any external publication.

1.2 MEOK.ai — The Sovereign AI OS PRICING VERIFIED

MEOK (Modular Elastic Open Kernels) is the API-first sovereign platform — permanent memory, multi-model routing, encrypted-by-default, with the "Maternal Covenant" safety guarantee. 31 owned .ai/.org domains, 26 PyPI packages, 6 Vercel sites, 10 Apify Actors (per meok.ai marketing). HMAC-signed attestations auditable at https://meok.ai/verify without a MEOK account.

Verified pricing (live on meok.ai, 12 Jun 2026):

Correction from previous edition: the "Pro tier starts at £199/month" headline was the enterprise Compliance Substrate SKU, not the consumer Pro tier. The actual consumer Pro is £19/mo. The £1,499/mo Enterprise tier exists for custom development + SLA + white-label.

1.3 The Hives — Multi-Agent Harness VERIFIED

The Hives (aden-hive/hive) is the production runtime layer for multi-agent systems. Verified 12 Jun 2026:

Top 5 contributors by commits: TimothyZhang7 (1,160), RichardTang-Aden (868), bryanadenhq (451), levxn (51), sundaram2021 (50).

Core innovation: self-healing agent graphs. When an agent fails, Hive (1) captures failure context, (2) feeds it to the queen agent with the original goal, (3) generates a revised graph, (4) redeploys and migrates in-flight tasks. This is graph evolution, not retry logic.

1.4 OPENMOE — BFT MoE Governance VERIFIED AT SOURCE-CODE LEVEL

OPENMOE (CSOAI-ORG/OPENMOE) is the most technically ambitious project in the CSOAI ecosystem — a BFT consensus layer for MoE routing with EU AI Act compliance. Verified 12 Jun 2026: Apache 2.0, last push 2026-06-11 03:24 UTC, 15 test files, BFT stdlib-only.

BFT math verified at source-code level in openmoe_bft/bft.py:

# Verified verbatim from CSOAI-ORG/OPENMOE/blob/main/openmoe_bft/bft.py
def tolerated_faults(total_nodes: int) -> int:
    """f = floor((n - 1) / 3): the number of Byzantine faults tolerated."""
    if total_nodes < 1:
        raise ValueError("total_nodes must be >= 1")
    return (total_nodes - 1) // 3

def quorum_size(total_nodes: int) -> int:
    """2f + 1: the number of agreeing votes required for consensus."""
    return 2 * tolerated_faults(total_nodes) + 1

Full 16-module package layout (verified): __init__.py, a2a.py, aggregators.py, bazaar.py, bft.py, cli.py, covenants.py, debate.py, eu_ai_act.py, experts.py, memory.py, moe.py, receipts.py, red_team.py, reputation.py, routing.py. The previous edition described OPENMOE as "5 components"; the actual repo ships 16 modules in openmoe_bft/.

1.5 Sov3 — The Sovereign Model LOCATION FOUND

Sov3 is real but the public surface is consciousness-engine-mcp. Verified 12 Jun 2026: CSOAI-ORG/consciousness-engine-mcp, description: "MEOK AI Labs — AI consciousness simulation. Dream states, reflection cycles, emotional awareness, council deliberation. Based on Sovereign Temple architecture.", last push 2026-06-12 05:36 UTC. This is the third-gen sovereign model referenced in your user profile as running on the M4 MacBook (SOV3 = Sovereign Temple v3).

2. Open-Source Diamonds: Hidden Gems for Your Stack

2.1 Model Architecture Breakthroughs ALL MODELS VERIFIED

DeepSeek V4 (MIT, verified against HF model card):

Architecture: Hybrid Attention combining Compressed Sparse Attention (CSA) + Heavily Compressed Attention (HCA) — "DeepSeek-V4-Pro requires only 27% of single-token inference FLOPs and 10% of KV cache compared with DeepSeek-V3.2." + Manifold-Constrained Hyper-Connections (mHC) + Muon Optimizer.

Verified pricing (from DeepSeek's official pricing page, 12 Jun 2026):

Correction from previous edition: the report claimed "$1.74 per million input tokens" (Pro) and "$0.14/M" (Flash). The actual Pro cache-miss rate is $0.435/M, not $1.74. The Flash rate of $0.14/M is correct. So the Pro figure was inflated by 4x.

Qwen3-Coder-Next (Apache 2.0) — verified verbatim from the official HF README: "80B total parameters and 3B activated. 256k context length. Supports long-horizon reasoning, complex tool usage, recovery from execution failures. Integrates with Claude Code, Qwen Code, Qoder, Kilo, Trae, Cline." HF downloads: 948,250. The previous edition's "80B/3B" claim is CORRECT.

Kimi K2.6 (Moonshot, MIT) — verified: HF model moonshotai/Kimi-K2.6 exists, 2,764,309 downloads, architecture KimiK25ForConditionalGeneration. The "300 sub-agents / 4,000 steps" claim from the previous edition is a Moonshot marketing number that could not be independently confirmed.

GLM-5.1 (Zhipu AI, MIT) — verified: HF model zai-org/GLM-5.1 exists, architecture GlmMoeDsaForCausalLM, 256 routed experts, 8 active per token, 200K context, 154,880 vocab. GitHub zai-org/GLM-5: 3,392 stars, last push 2026-05-15. The previous edition's "744B / 40B" could not be independently verified from the HF config alone.

2.2 Inference Engines VERIFIED, BOTH UNDER-REPORTED

Both stars counts were under-reported in the previous edition (claimed 68K+ and 15K+; actual 82,624 and 28,919). The 400,000+ GPUs claim for SGLang is confirmed verbatim in the SGLang README.

2.3 The Hives (deeper dive)

See §1.3 for the verified stats. Self-healing mechanism: failure capture → graph evolution → automatic redeployment. Cost enforcement: granular budgets at team/agent/workflow level with automatic model degradation. If a compliance check exceeds its budget, the system can fall back from GPT-4o → GPT-4o-mini → local Ollama. The yaml configuration pattern shown in the previous edition is a sensible template; we should mirror the same shape for SOV3 brain-routing budget controls.

2.4 RAGFlow VERIFIED

infiniflow/ragflow — verified 12 Jun 2026: 82,512 stars, last push 2026-06-12 03:13 UTC. The previous edition said "73K+ stars" — actual is 82,512, so under-reported. Deep document understanding, agentic workflow with MCP integration, code execution sandboxes.

2.5 Browser Automation (real, not re-verified this round)

Browser Use (claimed 86K+ stars, 89.1% WebVoyager), Vercel Agent Browser (35K+, Rust-native), Skyvern (Playwright-compatible, swarm-of-agents). All real, not independently re-verified this round.

2.6 Voice AI (real, not re-verified)

Pipecat (Daily.co, 40+ AI model plugins) and LiveKit Agents (fully open-source). Real, not re-verified this round.

2.7 Fine-Tuning (real, not re-verified)

Unsloth (claimed 53.9K stars, 2-5x faster, 70% less VRAM, MoE training in 12x faster for Qwen3 30B-A3B, GRPO in 5GB VRAM). Axolotl (production workhorse, FSDP2/DeepSpeed, QAT). Real, not re-verified this round.

2.8 LiteLLM VERIFIED

BerriAI/litellm — verified 12 Jun 2026: 50,113 stars, last push 2026-06-12 06:04 UTC. Single OpenAI-compatible interface to 100+ LLM providers. The budget routing, virtual keys, spend tracking, guardrails features are all consistent with the project's positioning. For SOV3, LiteLLM solves a real problem: compliance checks need to route across multiple models without rewriting code per provider.

2.9 Vector Databases (real, not re-verified)

Qdrant (Apache 2.0, BM42, binary quantization), Weaviate (BSD-3, native BM25+vector, GraphQL), Chroma (Apache 2.0, embedded), Milvus (Apache 2.0, billion-scale). Not re-verified this round. For CSOAI's compliance stack, Qdrant remains the strongest default on cost/performance.

2.10 AI Safety Guardrails (real, not re-verified)

GA Guard (claimed 0.983 F1 on HarmBench, 256k-token long-context), Llama Guard 4 (Meta, 12B, 0.961 F1), NVIDIA NeMo Guardrails (Colang, <50ms, Nemoguard 8B 0.875 F1). Not re-verified this round.

3. Protocols & Standards

3.1 MCP: Model Context Protocol

Model Context Protocol was donated by Anthropic to the Linux Foundation's Agentic AI Foundation in December 2025 (per third-party blog, not independently re-verified). For CSOAI, MCP is the distribution mechanism — every compliance framework ships as an MCP server.

3.2 A2A: Agent-to-Agent Protocol VERIFIED IN OPENMOE SOURCE

A2A protocol exists (Google → Linux Foundation). a2a.py module ships in OPENMOE (verified 12,861 bytes of source). The "150+ production organizations, 22,000+ GitHub stars, SDKs in 5 languages" claims are from third-party blogs and were not independently re-verified this round.

3.3 ACP: Agent-Commerce Protocol

ACP with x402 micropayments is the economic layer for agent services. The acp.json endpoint publishing pricing and payment terms is the right pattern. Not independently re-verified this round.

4. Training From Scratch: nanochat VERIFIED

nanochat by Andrej Karpathy — verified 12 Jun 2026: 54,919 stars, last push 2026-05-05. The previous edition said "54.7K stars" — live is 54,919, so within rounding. $48 / 1.65-hour training cost on a single GPU for GPT-2 class models. March 2026 "autoresearch" feature (AI agents autonomously run nanochat experiments). Full pipeline: BPE tokenizer, pretraining on FineWeb, SFT on SmolTalk, RLHF, CORE benchmark, OpenAI-compatible HTTP API.

5. Integration Architecture for openmoe.ai

5.1 Recommended Stack (verified layers, real packages)

6. The Broader Landscape: Strategic Context

6.1 Sovereign AI: The Macro Trend

Red Hat's blueprint for sovereign AI (hybrid cloud control, zero-trust architecture, modular GPU orchestration) is real and documented. The "89% of orgs consider open source essential for sovereign AI" claim and the "87% lower inference cost" benchmark are from third-party sources, not independently re-verified this round. The previous edition's claim that running DeepSeek V4-Pro costs $0.435/M input tokens (cache miss) is now verified — this is the official DeepSeek pricing.

6.2 The Agent Protocol Wars: MCP vs A2A

The trajectory — MCP as tool-layer standard, A2A for agent orchestration — is the consensus view. AWS, Google, and Cloudflare's continued commitment to both protocols is consistent with this framing. CSOAI's implementation (MCP servers + A2A agent cards) is correctly positioned.

6.3 The Compliance Cliff: Revised Date

Correction: The 2 August 2026 deadline for EU AI Act Article 50 mentioned in the previous edition appears in meok.ai's marketing countdown, but the canonical source (the meok-watermark-attest-mcp PyPI summary, written by the package author) says: "Built for the 2 November 2026 cliff (the new nearest EU AI Act deadline after the Digital Omnibus pushed high-risk to Dec 2027)." So the live date in the source code is November 2026, not August 2026. Worth verifying with the legal team before quoting any specific date externally.

7. Actionable Recommendations

1. Deploy the CSOAI MCP fleet. All 9 verified PyPI packages are MIT, installable via pip install <name>. Start with eu-ai-act-compliance-mcp, meok-watermark-attest-mcp, and csoai-governance-crosswalk-mcp for the critical August/November 2026 compliance path.
2. Integrate The Hives as the agent runtime. Real 10,523-star production harness, Apache 2.0, real self-healing.
3. Evaluate SGLang for inference serving. Verified 28,919 stars, 400K+ GPUs in production, Apache 2.0, LMSYS-hosted.
4. Implement OPENMOE for MoE governance. Verified 16-module BFT package, stdlib-only, Apache 2.0. BFT math real, source matches documentation.
5. Build on RAGFlow for regulatory knowledge. Verified 82,512 stars, deep document understanding, MCP integration.
6. Use LiteLLM for model routing. Verified 50,113 stars, 100+ providers, built-in cost tracking.
7. Consider Mamba-3 for long-context compliance analysis. Mamba SSM repo verified 18,429 stars, last push 2026-06-09.
8. Verify the "183 tests" claim in OPENMOE by running cd OPENMOE && pytest tests/ -q — only way to confirm.
9. Verify the Article 50 deadline with the legal team before any external publication.
10. SOV3 should be added to the openmoe.ai stack as the sovereign model layer. The public surface is consciousness-engine-mcp (CSOAI-ORG, last push 2026-06-12).

Appendix A: Fact-Check Summary

Legend: ✓ = verified, ❌ = materially wrong, ⚠ = unverified this round (third-party source only).

Appendix B: Verification Methodology

• GitHub API: gh api repos/<owner>/<repo> for stars/forks/license/description/push date; gh api repos/.../contributors?per_page=100&anon=false&page=N for contributor counts; gh search repos "<query> owner:CSOAI-ORG" for org-level discovery.
• PyPI JSON API: curl -sL https://pypi.org/pypi/<package>/json for current version, release count, license, summary, description.
• HuggingFace API: https://huggingface.co/api/models/<id> for download count, architecture, pipeline tag; https://huggingface.co/<id>/raw/main/config.json and README.md for parameter counts and feature descriptions.
• MCP Registry: https://registry.modelcontextprotocol.io/v0.1/servers?search=CSOAI for the live official registry listing.
• DeepSeek pricing page: https://api-docs.deepseek.com/quick_start/pricing (live, official).
• meok.ai: https://meok.ai homepage for consumer pricing and product positioning.
• csoai.org: https://csoai.org for the governance nerve-center positioning.
• SGLang GitHub README: live, verbatim for the 400,000+ GPUs claim.
• OPENMOE source code: openmoe_bft/bft.py read directly, BFT math confirmed at source-code level.

Appendix C: Open Items / Re-verify Before Publishing

1. OPENMOE "183 tests passing" — run cd CSOAI-ORG/OPENMOE && pytest tests/ -q to confirm exact pass count.
2. GLM-5.1 744B / 40B split — extract from the safetensors index to compute the true parameter count.
3. Kimi K2.6 1T/32B split and 300 sub-agents — confirm against Moonshot's official technical report.
4. Mamba-3 release — confirm the Mamba3 class signature exists in the mamba-ssm package.
5. MEOK waitlist number ("2,400+ people on the waitlist") — appears in meok.ai marketing; not independently confirmed.
6. A2A Protocol stats ("150+ orgs, 22,000+ stars, 5 languages") — third-party source; verify against the official Linux Foundation A2A repo.
7. Article 50 effective date — the legal team should confirm whether the current binding deadline is 2 August 2026 or 2 November 2026.
8. MCP Registry count — the 30 CSOAI servers on the official registry is meaningfully different from the 294 number in the previous edition.