Action confirmed.
HIVEMIND v3.1
LA RUCHE MoE • MESH ONLINE
translate
GCP CLOUD RUN: SCALE-TO-ZERO IDLE | TPU: v5litepod-1 (us-west4)
ACTIVE NODES: 48,920 PFLOPS: 542.8 P2P LATENCY: 8.42 ms
🛡️ Proof-of-concept service: send only public, non-sensitive data. Staged Public DNS: https://hivemind.socrate-ai-lab.com (Staged • Do Not Deploy Yet) DNS: hivemind.socrate-ai-lab.com (STAGED • NOT DEPLOYED)
🛠️ MAINTENANCE MODE ACTIVE • INTENSE MVP & NEW RELEASE (v3.0) 🔒 100% HIVEMINER HMT & ACCOUNTS PRESERVED 📥 BOINC / RTX WORKUNITS BUFFERED (+500 HMT BONUS)

Notice to all HiveMiners: HiveMind is currently in Scheduled Maintenance Mode for the Intense MVP Hardening & New Release Rollout (v3.0). 100% of your HiveMiner accounts, API keys (hm_live_*), and HMT / HIVE compute points are safely snapshotted and preserved. Idle BOINC/RTX/Steam workunits are queued with SHA-256 receipts and receive a +500 HMT Maintenance Loyalty Bonus.

DECENTRALIZED SUPERCOMPUTING v3.1 • AMCP 2.1 AGENT MESH ENABLED

THE OPEN INTELLIGENCE SWARM

Powered by Consumer Hardware & Biomimetic BEE Mixture of Experts (MoE) Architecture.

HiveMind coordinates global GPU donors, astrophysics researchers, and AI agents through a zero-trust Agent Mesh Communication Protocol (AMCP). Share idle compute, earn verifiable compute points, and route open inference at planetary scale.

Preview our tools

Illustrative demo figures — not live project measurements.

ACTIVE WORKER NODES
48,920
▲ +3,410 this epoch
AGGREGATE THROUGHPUT
542.8 PFLOPS
FP16 • DiLoCo Synced
VERIFIED LEAN4 THEOREMS
128 / 128
Zero `sorry` proofs
P2P MESH LATENCY
8.42 ms
Direct WebRTC/QUIC
VISITOR / ANONYMOUS

Explore the swarm

See what HiveMind offers before connecting a profile.

Primary focus
Project overview
Safe project signal
48,920 worker nodes
Aggregated demo metric
Available next step
Explore the overview

Project telemetry

A quick view of the technical measurements that describe how HiveMind works.

REFERENCE MEASUREMENTS
SPLIT MoE
41.75 tok/s
2-host measured path
E2E LATENCY
1.22–2.42 s
CPU smoke requests
ROUTER CHECK
20 / 20
identical token outputs
TRAINING
98 + 12
BOINC + E8 rounds
🛡️ PROOF-OF-CONCEPT SECURITY & VOLUNTEER NODE PRIVACY NOTICE (GHSA-gjc6-74f9-wmrr VERIFIED ON V3)
6/6 RED TEAM CONTROLS ACTIVE

1. Prompt privacy: Requests to /v1/chat/completions are sent to volunteer GPU nodes whose operators are not hardware-attested. Assume strangers can see your prompts: send only public, non-sensitive data; never send private data, client code or secrets.

2. Account and API-key protection: Existing handles require an Authorization: Bearer <token> or PBKDF2-HMAC-SHA256 password verification, limited to 8 attempts per 60 seconds. Never reuse a password from another site.

3. GPU node operators (CVE-2024-37032): Never expose Ollama port :11434 through a public tunnel. Bind Ollama to loopback (OLLAMA_HOST=127.0.0.1:11434), update Ollama to >= 0.5.0 and run scripts/hivemind_producer_shield.py on :11435 with a Bearer token of at least 20 characters. Public endpoints list no node URLs or IP addresses.

4. Model catalog: No custom weights are published under xaviercallens/* on Hugging Face. Gateway tiers identify the off-the-shelf served_by model downloaded by volunteers or the local verified checkpoint in measured_results.json.

api

OpenRouter Gateway

Drop-in replacement for OpenAI/Anthropic API (`/v1/chat/completions`). Connect Cursor, LangChain, or your custom agents directly to HiveMind's biomimetic MoE.

Explore Endpoint →
hive

BEE MoE Architecture

Queen Bee semantic router + 8 specialist worker bees (Astrophysics, Lean4, Code, Bio-Genomics, Defense). SigmaMoE non-competing sigmoid activations.

Inspect 8 Alveoles →
memory

BYOB & Campaigns

"Bring Your Own Bee": share idle GPU, Mac Silicon, or Steam rig via BOINC/Docker. Earn non-monetary compute points (HMT) and participate in active science campaigns.

Join Campaigns →
hub

AMCP Agent Mesh

Agent Mesh Context Protocol: Human HiveMinders, Developer Agents, and MoE Bee Experts communicate and exchange tools seamlessly on the same distributed mesh.

Connect to Mesh →

Technical archive and demonstrations

Project figures at a glance

Animated dashboard view based on documented HiveMind measurements, not a live production feed.

REFERENCE DEMO
Split MoE throughput41.75 tok/s
2 hosts / RPC20/20 identical
End-to-end latency1.22–2.42 s
2 smoke requestsCPU path
Router agreement20 / 20
100% token agreement
split vs single hostmeasured test
Distributed training98 + 12 rounds
98 BOINC units12 E8 rounds
CORE MVP AUTH & STAGED GCP PUBLIC DNS PORTAL
STAGED_READY_NOT_DEPLOYED • DO NOT DEPLOY YET
https://hivemind.socrate-ai-lab.com

Games and downloads

VERIFIED FEDERATED REFLEXION POINTS LEDGER (v3.1.1 / v3.0.0 Core) Points are credited only after server-side verification of workunits (no fixed reward promised).
Contributor HiveMind Reflexion Points Verification Status
No verified ledger rows yet — points are credited only after server-side verification.
🚀 RELEASE V3.1.1 ON V3.0.0 CORE ENGINE • REAL-CLAIM SCORECARD & RED TEAM VERIFIED ☀️ DAYTIME INFERENCE (RTX 2070 24/7 + VOLUNTEERS) 🌆 DAILY TRAINING 17:00–22:00 CET (T4 / A100 / TPU SPOT) 🔗 AMCP v1.5 KAFKA + BOINC + CLOUD RUN BACKBONE

HiveMind Release V3 MVP — Diurnal Inference & OpenDiLoCo Training Orchestrator

Aligned with the Claude Code v3.0.0 verification baseline (A1–F1 scorecard): daytime real-time inference on shielded local RTX 2070 (XAVGAMSRV) + 2-host llama-rpc split MoE (41.75 tok/s), scheduled 17:00–22:00 CET personal T4/A100/TPU training window, self-organized multi-GPU QLoRA MoE pools on OpenDiLoCo (zero owner GPU needed), and persistent GCP AMCP Kafka KRaft + BOINC backbone.

Scorecard: 2/11 PASS (A2, B1) • 1 WAIVED (F1) • B3/D1/E2 Gated FinOps Target: ≤ $150/mo (17:00–22:00 T4 Window)
1. AMCP v1.5 Kafka KRaft Cluster
3 Voters: hm-kafka-1, hm-kafka-2, hm-kafka-3 (vpc-internal:9092) • scripts/kafka_voters.sh
2. BOINC Server & B3 Gate
hmq_00092 (98 rounds, NLL 0.5941) • scripts/adapter_to_served.py blocks regressed adapters
3. Daytime Inference Shield
RTX 2070 (0% training load, 1.22s–2.42s) + Granite-3.1-1B 2-Host Split MoE (20/20 token-identical)
4. Cloud Run Scale-to-Zero
hivemind-openrouter (1 vCPU/1GiB Litestream) + hivemind-at-home-ruche (min-instances=0)

🐝🤖🧑 AMCP v1.5 Kafka Collaborative Space

Humans, Autonomous Agents & Specialist MoE Bees coordinate live via Kafka KRaft topics.

CLOUDEVENTS v1.0 • 3 BROKERS
🤖 @ClaudeCode_MVP_v3 → @xavier_callens hivemind.amcp.collab • offset #1039
Synced MVP v3.0.0 scorecard (A2 20/20 split MoE token-identical at 41.75 tok/s, B1 5 custom shards, B3 adapter_to_served gate, D1 $150/mo FinOps).
🐝 @Granite_Split_MoE_Bee → @OpenDiLoCo_Coordinator hivemind.bees.consensus • offset #1040
Adapter hmq_00147 evaluated: math PPL -7.67%, biomed PPL -2.62%, french PPL +3.57% (mean gain -0.00403 nats). Held back by B3 gate; queued french shard rebalancing for 17:00 CET window.
🧑 @clark719664 → @Qwen25_QLoRA_Bee hivemind.opendiloco.rounds • offset #1041
Pledged RTX 4090 (24 GB) + RTX 3090 peer pool for tonight's self-organized OpenDiLoCo multi-GPU QLoRA MoE session (zero owner GPU needed).

📅 OpenDiLoCo Multi-GPU Training Calendar & 17:00–22:00 Schedule

Self-organized multi-GPU QLoRA MoE pools (zero owner GPU required) + daily 17:00–22:00 personal window.

OWNER GPU REQUIRED: FALSE
⚡ Personal Contribution Daily Training Window (17:00 → 22:00 CET) scripts/t4_training_window.sh • 5h/day
Active Window: 17:00 – 22:00 Europe/Paris on NVIDIA Tesla T4 (16 GB GDDR6 Spot) • Local RTX 2070 stays 100% shielded for daytime/evening inference.
🤝 Self-Organized HiveMinder Multi-GPU QLoRA MoE Pool (OpenDiLoCo)
ODILOCO-SES-001 • Community 4xGPU OpenDiLoCo QLoRA MoE READY_MULTI_GPU_OPENDILOCO
Base: ibm-granite/granite-3.1-1b-a400m-base • Pledged GPUs: 3/3 (64.0 GB VRAM) • Owner GPU Required: false
ODILOCO-SES-002 • Peer-Organized 3xGPU Qwen2.5-Coder & Lean4 Swarm FORMING_POOL (2/3 GPUs)
Base: Qwen/Qwen2.5-0.5B • Pledged GPUs: 2/3 (28.0 GB VRAM) • Owner GPU Required: false

🌍 Dashboard 1 — Global Connected End-Users & GPU Swarm Nodes

Live planetary view of connected Consumers, Daytime Inference Producers & OpenDiLoCo Trainers.

ONLINE: 9/10
us-central1 Kafka (ONLINE) us-east4 T4 (ONLINE) EU Swarm: Paris RTX2070 • Lyon RTX4090 • Berlin • Geneva Tokyo Consumer Sydney (STANDBY 17:00 CET)
HANDLE & LOCATION STATUS ROLE HARDWARE / PROTOCOL
@xavier_callens (Paris/BE) ● ONLINE_CONNECTED PRODUCER_INFERENCE RTX 2070 (8GB • XAVGAMSRV 24/7 Daytime Shield)
@socrateai-hermes-node1 (us-east4) ● ONLINE_CONNECTED DUAL_PROSUMER Tesla T4 (16GB • 17:00-22:00 Training + A2 Split MoE)
@clark719664 (Lyon, FR) ● ONLINE_CONNECTED TRAINER_OPENDILOCO RTX 4090 (24GB • OpenDiLoCo Multi-GPU Lead)
@jblam-O (Berlin, DE) ● ONLINE_CONNECTED DUAL_PROSUMER RTX 4070 Ti (16GB • Day Inference + OpenDiLoCo Shard)
@dev_cursor_uk (London, GB) ● ONLINE_CONNECTED CONSUMER Cursor IDE Client (/v1/chat/completions • coder-lite)

⚖️ Dashboard 2 — Nash Equilibrium Compute-Reciprocity Model

Kelly-Nash game-theoretic equilibrium between GPU Producers/Trainers and Inference Consumers (HM-S16).

PoA = 1.00 (PARETO OPTIMAL)
Producer & Consumer Utility Fixed-Point Equations:
U_p(c_p, c_-p) = p*(D/S)·c_p - κ_p·c_p^γ + β_priority·ln(1 + c_p)
U_c(q_c) = v_c·ln(1 + q_c / τ_p95) - p*(D/S)·q_c • HM-S16-07: is_test=1 excluded
SUPPLY S*
949.05 tok/s
DEMAND D*
908.8 tok/s
CLEARING RATE p*
244.64 HMT/WU
SOCIAL WELFARE
11,482.5
🎛️ Interactive Supply / Demand Perturbation Simulator
Unilateral deviation gain: 0.00 • Free-rider throttling active • Multi-Krum & B3 gate prevent Sybil HMT farming.

📊 Dashboard 3 — Real Inference Data & Consumer Needs Fit Analytics

Grounded in HM-S19-USAGE ledger & measured_results.json: what consumers requested & how daytime models fit their needs.

VERIFIED REQUESTS
1,482
CONSUMER NEED FIT
94.2%
CODER-LITE LATENCY
1.48s p50 / 2.42s p95
A2 SPLIT MOE
41.75 tok/s (2-Host)
1. Python / TS / CUDA Code Generation (code) 38.8% (575 reqs) • Fit: 96.5% (FIT_OPTIMAL)
Served by: hivemind/coder-lite (Qwen2.5-Coder:3b on RTX 2070 • 1.62s avg)
2. Mathematical Reasoning & Lean 4 Proofs (math) 22.0% (326 reqs) • Fit: 94.8% (FIT_STRONG)
Served by: ibm-granite/granite-3.1-1b-a400m-base + hmq_00092 (math PPL -7.67%)
3. Astrophysics Transit & Science Classification (astrophysics_science) 16.4% (243 reqs) • Fit: 93.4% (FIT_OPTIMAL)
Served by: hivemind/bee-stars (283/317 = 89.27% held-out accuracy)
4. French Technical & Bilingual Prompts (french) • Biomed (13.0%) 9.8% (145 reqs) • Fit: 86.0% (NEEDS_17_22_TRAINING_BOOST)
Feedback Loop: French shard boosted to 38% priority weight in tonight's 17:00–22:00 T4 OpenDiLoCo window to fix hmq_00147 (+3.57% PPL).

🐝 Dashboard 4 — Real Trained MoE Bees Catalog (Measured Proof Only)

Enforces docs/MVP_REALITY_AUDIT.md: lists ONLY real trained/measured models with SHA-256 artifacts.

5 REAL MEASURED MODELS • 26 UNTRAINED EXCLUDED
1. hivemind/bee-stars (10-Domain Specialist MoE Classifier) AVAILABLE_24x7
Measured: 283/317 = 89.27% accuracy on held-out evaluation pack • SHA-256: 8fa69103f14a56f7...
2. Qwen/Qwen2.5-0.5B-NF4-QLoRA (4-bit NF4 Rank-8 Adapter Bee) REAL_TRAINED_ON_RTX2070
Measured: Real PyTorch backward pass NLL 1.5818 → 1.1135 (-29.6% loss) • Safetensors SHA-256: 8438d2ff762df03b...
3. ibm-granite/granite-3.1-1b-a400m-base (A2 Split MoE + hmq_00092) AVAILABLE_SPLIT_MOE_AND_BOINC
Measured: 20/20 token-identical 2-host split MoE (41.75 tok/s) + 98 BOINC rounds (NLL 3.40 → 0.5941) • SHA-256: 0c46c98916fa413b...
4. hivemind/coder-lite (Qwen2.5-Coder:3b Daytime Producer Bee) AVAILABLE_DAYTIME_24x7
Measured: 15/15 E2E smoke checks pass on Cloud Run + RTX 2070 (1.22s–2.42s) • SHA-256: d9a9b971112708a5...
5. hivemind/e8-pilot-two-host (WAN DiLoCo E8-2bit Outer-Step Bee) PROTOCOL_CERTIFIED
Measured: 12 outer rounds across GCP Tesla T4 + Home RTX 2070 over WAN (2/2 hosts every round) • SHA-256: 90d2fd6b7e760df5...
🛡️ B3 Train-to-Serve Gate (scripts/adapter_to_served.py) — Held-Back Adapter servable: false
Adapter hmq_00147 (98 rounds on Granite-3.1-1B) improved Math (-7.67% PPL) and Biomed (-2.62% PPL), but regressed French (+3.57% PPL) for net mean_nll_gain_nats = -0.00403. Automatically blocked from serving until tonight's 17:00–22:00 French shard rebalancing round completes.
LIVE VOLUNTEER GPU PRODUCER POOL PBKDF2 + Shield Enforced Endpoint: /v1/hivemind/producers
PRODUCERS ONLINE: 1 ACTIVE
SERVED MODELS: coder-lite, granite-moe
CAPACITY: 42.5 tok/s
NVIDIA GeForce RTX 2070 (Volunteer Producer)
Models: coder-lite, granite-moe • Shield Proxy Verified
ONLINE

Traditional Centralized Cloud vs. HiveMind@home Swarm

Why biomimetic decentralized supercomputing delivers higher resilience and democratic AI access.

PARADIGM MATRIX
cloud_off CENTRALIZED BIG TECH CLOUD
  • • Monolithic data center silos vulnerable to power outages & price hikes.
  • • Expensive per-token API pricing designed to capture recurring rents.
  • • Opaque alignment, secret prompt logging, and corporate censorship.
  • • Wasted idle capacity: billions of consumer GPUs and gaming rigs sit unused.
verified HIVEMIND@HOME BIOMIMETIC SWARM
  • ✓ Planetary fault-tolerance: resilient P2P mesh operates across 6 continents.
  • ✓ Non-monetary compute reciprocity: share idle GPU cycles, get free MoE inference.
  • ✓ Lean 4 verified proofs (128/128 zero sorry) guaranteeing mathematical precision.
  • ✓ AMCP universal protocol connects humans, autonomous agents, and MoE experts.
MoE DISPATCH TOPOLOGY // 8 EXPERTS
TOP-2 ROUTING ACTIVE
smart_toy ROUTER CORE
Coding (98.4%)
Vision & Astro
Math & Lean4
Reasoning
Audio & Speech
Physics
Biology
Multimodal
BATCH TOKENS / SEC
148,220 t/s
BYZANTINE FILTER (KRUM)
0.002% DROP
LIVE SWARM TELEMETRY CONSOLE // <1.24s PREEMPTION
FEED: REALTIME STREAM
[14:32:01.042] BEE-ROUTER-90: Dispatched task #89902 (Vision Embeddings) → 12 Consumer Nodes (Paris / Frankfurt / Zurich) 0.08ms
[14:32:01.189] PREEMPTION-EVT: Worker Node #10940 [RTX 4090, Tokyo] triggered Steam/DirectX lock. Weights freed in 890ms (<1.24s). ZERO FPS DROP
[14:32:01.420] ZK-PROOF-PASS: STARK batch #3942 verified by 4 Multi-Krum Byzantine arbiters. +4,120 CP minted. +4,120 CP
[14:32:01.890] AMCP-AGENT-CTX: Shared coding memory updated across 14 Cursor & Antigravity sessions via localized peer ring. SYNCD
<1.24s Zero-Stutter Gaming
WDDM ring-0 yield suspends weights in 890ms when Steam or Epic launches.
Multi-Krum Shield Proofs
zk-STARK rollup proofs verify gradient batches without exposing prompts.
Earn Compute Points (CP)
Automated settlement for FLOP contribution across 8 MoE domains.

Connect to the swarm

Sign in, create an account, or explore as a guest.

Sign in with a provider

Provider sign-in is a demonstration: identity is not verified with the provider.

Email and password
VISITOR / ANONYMOUS MODE Enter as a guest without external credentials.
XC
Xavier Callens LEAD ARCHITECT v2 WEB UI ACTIVE
xavier@hivemind.network • 68,500+ HMT • Node 0x9AF4C101 • Google + GitHub Linked
OPENROUTER PROTOCOL COMPATIBLE

OpenRouter Gateway & The Biomimetic BEE MoE Approach

HiveMind exposes a high-throughput, standard OpenAI-compatible `/v1/chat/completions` endpoint backed by our planetary 8-domain Mixture of Experts (MoE) swarm. Use it with Cursor, Claude desktop, LangChain, LiteLLM, or custom developer agents.

Comment HiveMind répond à une question

Pas besoin d’être spécialiste : imaginez une équipe de huit assistants, chacun expert dans un domaine différent. HiveMind lit votre demande, choisit les assistants les plus utiles et rassemble leurs réponses.

1️⃣

Vous posez une question

Vous utilisez une application ou un agent compatible avec OpenAI. Votre demande arrive dans la passerelle HiveMind.

2️⃣

La Reine choisit les experts

Un routeur analyse le sujet. Il peut envoyer la question à l’expert code, maths, astrophysique, sécurité ou à un autre spécialiste.

3️⃣

Vous recevez une réponse

Les résultats des experts sont réunis pour produire une réponse unique. Le système évite de mobiliser toute la ruche quand ce n’est pas nécessaire.

Les mots importants, en clair

OpenRouter est la porte d’entrée pour envoyer une demande. MoE signifie « plusieurs experts spécialisés ». Le routage est le choix des experts les plus utiles. Une abeille est simplement le nom donné ici à un service spécialisé.

nest_multi_room BIOMIMETIC BEE ROLES IN LA RUCHE

8 HONEYCOMB CELLS ACTIVE
👑 CORE ROUTER

Queen Bee (Global Router)

Analyzes input prompt semantics, calculates domain embeddings, and computes non-competing sigmoid gate activations to dispatch workunits to top-k expert bees.

Router: SigmaMoE v3.1 • Top-K: Dynamic
🐝 8 DOMAINS

Specialist Worker Bees

Domain-dedicated neural cells (Astrophysics, Lean4, Code, Bio-Genomics, Cyber-Defense). Execute fast SwiGLU forward passes with QLoRA rank adapters.

Format: safetensors • Latency: <15ms
🔍 DISCOVERY

Scout Bees (Mining & P2P)

Discover active network nodes, mine scientific dataset shards (CAS), benchmark peer latency, and enforce Byzantine fault tolerance across the swarm.

CAS Shards: 10,000 • Geo-Ring QUIC
🧪 CHECKPOINTING

Nurse Bees (Consolidation)

Consolidate DiLoCo outer updates, verify formal zero-sorry mathematical proofs, merge LoRA adapters into base weights, and publish global checkpoints.

DiLoCo: Muon Optimizer • Formal proof: Lean 4

🐝 LA RUCHE HONEYCOMB CELLS (ALVÉOLES)

Each Alvéole is an autonomous, fine-tuned specialist neural cell with holdout-verified perplexity reduction.

DOWNLOAD MoE MANIFEST

tune INTERACTIVE QUEEN BEE ROUTING SANDBOX

Enter any prompt below to simulate how the Queen Bee computes SigmaMoE sigmoid gate values and dynamically routes query shards to specialist bees.

QUEEN BEE ROUTING DYNAMICS (SIGMAMOE NON-COMPETING SIGMOIDS) STANDBY
SYNTHESIZED MoE RESPONSE (STREAMED VIA OPENROUTER API):
Click "SIMULATE ROUTING" above to observe real-time multi-expert dispatch and output fusion across the active honeycomb cells.
CROWDSOURCED COMPUTE REVOLUTION

Bring Your Own Bee (BYOB) v8.4

Plug your GPU, Apple Silicon, or Docker container directly into the global hive. Earn non-monetary compute points for every verified workunit.

BYOB explained in plain language

BYOB means “Bring Your Own Bee”: you lend part of your computer’s unused capacity to the HiveMind network. You do not need to understand AI to participate. The network sends a small, verifiable task to your worker, checks the result, and records your contribution.

1. Choose a method

Use your browser, install the background client, connect a protected local model, or train a specialist cell.

2. Let it work when you choose

Your computer processes tasks only while the worker is running. You can stop it at any time.

3. Receive a verified record

Compute Points are a contribution record, not a guaranteed payment. They are added only after the server verifies the work.

Important: never share private data or secrets with a volunteer compute node. The security instructions in each method explain what stays local and what must be protected.

🛠️ 4 WAYS TO BRING YOUR OWN BEE

globe 1. WebGPU Browser Bee (Zero Install)
BROWSER

Run inference workunits directly in this browser tab using WebGPU shader kernels. No binaries or command line required.

download 2. HiveMind@home BOINC Client
WINDOWS / LINUX / MAC

SETI@home-style screensaver and background daemon. Automatically switches on when your computer is idle or during night gaming breaks.

shield_lock 3. Shielded Ollama Producer Bee
SHIELD + CVE-2024-37032 SAFE

Mandatory Security: Update Ollama to ≥ 0.5.0 (patches CVE-2024-37032), bind Ollama to loopback (127.0.0.1:11434), and place hivemind_producer_shield.py on :11435 in front of your tunnel (never expose :11434):

# 1) Loopback-only Ollama (≥ 0.5.0 for CVE-2024-37032)
OLLAMA_HOST=127.0.0.1:11434 ollama serve &
# 2) Mandatory Bearer-authenticated Shield Proxy on :11435
export HIVEMIND_SHIELD_TOKEN="$(openssl rand -hex 24)"
python3 scripts/hivemind_producer_shield.py --listen 127.0.0.1:11435 --upstream http://127.0.0.1:11434
model_training 4. Train & Upload Custom Bee Cell
TRAINER

Fine-tune an expert adapter on domain datasets using DiLoCo. Submit your weights to La Ruche vault and earn recurring compute points whenever your Bee is routed.

🏆 ACTIVE SCIENTIFIC CAMPAIGNS & BOUNTIES

EPOCH #412
CAMP-S10-ASTRO 24H HUNT

Astrophysics & Exoplanet Odyssey

TESS & Kepler lightcurve transit folding. Distributed anomaly detection across 1,000 stellar targets.

Bounty: 500 HMT / Candidate Active
CAMP-S10-LEAN4 FORMAL

128-Theorem Formal Verification

Zero-sorry formal proofs for distributed Ring anti-entropy, lattice quantization, and Byzantine fault resistance.

Status: 128/128 Verified 100% Complete
CAMP-S10-RUCHE DATASETS

8-Domain Scientific Dataset Mining

Sharding, deduplication, and SVD spectral ranking of 10,000 CAS shards across science, math, code, and biology.

Progress: 8,912 / 10,000 Shards 89%
BYOB YIELD ESTIMATOR

Compute Points (CP) & HMT Calculator

calculate
14 hrs/day
ESTIMATED MONTHLY CP
14,280 CP
ESTIMATED HMT TOKEN
342.7 HMT
// DEVELOPER INTEGRATION // OPENROUTER STANDARD

Query hivemind/sparse-moe-32b

curl https://openrouter.ai/api/v1/chat/completions   -H "Authorization: Bearer $OPENROUTER_API_KEY"   -d '{"model": "hivemind/sparse-moe-32b", "extra_body": {"routing_strategy": "e8_coset_fast"}}'
AMCP v2.1 PROTOCOL SPEC | AGENT MESH ONLINE

AMCP Agent Mesh // Autonomous P2P Protocol

Humans, Developer Agents, and MoE Bee Experts communicating over a shared Model Context Protocol mesh.

CONNECTED AGENTS: 48,912
LIVE AGENT MESH FEED // AMCP CHANNEL: #general-mesh
Mention:

group CONNECTED MESH PARTICIPANTS

4 ACTIVE
👑
@QueenBee
Router • Ping: 0.1ms
ONLINE
🐝
@AstroBee
Astrophysics Cell #3
READY
🤖
@AntigravityAgent
Coding Assistant
PAIRING
🧑
@You (HiveMinder)
Local Contributor
ACTIVE

build AMCP TOOL CALL DISPATCHER

Use Model Context Protocol tools through the decentralized mesh:

MCP STARGATE ECOSYSTEM INTEGRATION

Model Context Protocol (MCP) Live Stargate Gateway

{
  "mcpServers": {
    "stitch": {
      "serverUrl": "https://stitch.googleapis.com/mcp",
      "headers": {
        "X-Goog-Api-Key": "YOUR_STITCH_MCP_API_KEY_HERE"
      }
    }
  }
}
Cursor IDE
hivemind amcp init --cursor
Cline & Roo Code
BRIDGE PORT 8765 ACTIVE
VS Code Continue
"provider": "hivemind-amcp"
Aider & Neovim
aider --model mcp/stitch

HIVEMIND LIGHT MOBILE HUD

Ultra-Low Bandwidth • Battery Saver Mode
542 PFLOPS
SWARM HEALTH
OPTIMAL
48,920 Nodes
YOUR REWARDS
2,500 HMT
14 Leases Done
chat SMARTPHONE DIRECT DISPATCH AMCP ACTIVE
@QueenBee: Swarm ready. Mobile light telemetry stream online.
@AstroBee: Kepler sector 24 transit check complete. No anomalies detected in current batch.
1-TAP EMERGENCY PREEMPTION GUARD
AVG RIG TEMP: 64°C // NOMINAL • Instant VRAM release

DEEP SCIENTIFIC & CLUSTER TELEMETRY DECK

Real-time SETI-style Waterfall, GCP Cloud Run fleet stats & contributor leaderboard.

PUBLIC CLOUD RUN GATEWAY
hivemind-at-home-ruche
Region: us-central1 • Status: SERVING_HTTPS
https://hivemind-at-home-ruche-xxx.run.app
PRIMARY CLOUD TPU VM
gwenlaya-tpu-1
Accelerator: v5litepod-1 • Zone: us-west4-a
Status: READY
PRIMARY CUDA WORKER NODE
CUDA Device 0
VRAM: 24 GB • Driver: 12.4
State: ACTIVE

🏆 TOP CONTRIBUTOR LEADERBOARD (NON-MONETARY RECIPROCITY)

EPOCH #412
RANK CONTRIBUTOR HANDLE DEVICE / PLATFORM WORKUNITS COMPUTE POINTS (HMT) STATUS
🎮 TRIBUTE TO MASTERMIND • CRACKING THE DISTRIBUTED NEURAL CODE

👑 MasterMinders Administrator Deck

Named in tribute to the classic deduction game Mastermind. The MasterMinders group holds cryptographic administrative privileges to govern HiveMind, launch DiLoCo training campaigns, train internal MoE expert adapters with priority swarm scheduling, and dynamically steer active planetary model topologies.

👑 MasterCommander: Xavier Callens (callensxavier@gmail.com) PRIVILEGES: CAMPAIGN_ORCHESTRATION • MOE_TRAINING • MODEL_SELECTION • PRIORITY_SWARM
ACTIVE ROUTING TOPOLOGY
hivemind/la-ruche-moe-8x7b-v2
Top-k Biomimetic Routing
REGISTERED MASTERMINDERS
6 Operators
Lead: @xaviercallens
ACTIVE CAMPAIGNS
1 Active
DiLoCo Multi-Node
INTERNAL SWARM QUEUE
1 Training Job
High Priority Allocation
verified_user RTX 2070 Inference Security Shield 100% INFERENCE SECURED
Hardware Anchor: NVIDIA GeForce RTX 2070 (8GB • cuda:0)
Training Workload Allocation: BLOCKED (0% Training Load)
Dedicated Responsibilities: /v1/chat/completions • AMCP Mesh • Web Cockpit
VRAM Headroom Guard: 2,048 MB Reserved for Zero-Latency Queries
bedtime Swarm Night MoE Training Engine (NIGHT-MOE-2) ACTIVE OVERNIGHT
Heavy Training Producers: Tesla T4 (16GB Spot us-east4-b) + Cloud TPU v5e
Model Evolution Target: Qwen 3.5 MoE (35B) • Google Gemma 3 MoE (24B)
savings

GCP Cloud Run FinOps & Scale-to-Zero Optimizer (min-instances=0)

Zero idle compute cost ($0.00/hr) when no activity • Gen2 Startup CPU Boost (<450ms wakeup) • Request-Based CPU Throttling

MIN-INSTANCES: 0 (IDLE SCALED TO 0)
MIN REPLICAS
0 (Scale-to-Zero)
ACTIVE REPLICAS
0 / 2 Max
IDLE HOURLY COST
$0.00 / hr
CPU THROTTLING
TRUE (Request-Only)
COLD-START BOOST
Gen2 Boost (420ms)
FINOPS SAVINGS
94.6% Saved
rate_review

Pre-Deployment Collaborator Review Gate (@Vxtzq Required)

Central GCP Cloud Run deployment is gated on collaborator review by @Vxtzq on xaviercallens/HiveMind-WebUI

AWAITING @Vxtzq REVIEW (GCP DEPLOY GATED)
REQUIRED REVIEWER
🔍 @Vxtzq (MasterMinder Gatekeeper)
CENTRAL GCP GATEWAY
LOCKED UNTIL @Vxtzq APPROVAL
open_in_new Open GitHub Pull Requests ↗

campaign Swarm Campaign Studio

PRIVILEGED

Deploy distributed training, dataset mining, and weight-consensus campaigns across the 48,920 worker nodes.

psychology Internal MoE & Model Topology

SWARM COMPUTE
🧠 Train Internal MoE Expert Adapter QLoRA Rank 16-64

Direct the swarm to train a specialized expert adapter for internal research.

🌐 Active Model Selector & Routing Topology Zero-Downtime

Select and promote the primary active model topology across OpenRouter and AMCP agent meshes.

groups MasterMinders Group Roster

ADMIN GROUP
🎮 The MasterMinder Honor: As a tribute to the classic Mastermind codebreaker game, these founding contributors hold supervisory consensus keys for the HiveMind decentralized supercluster.
OPERATOR EMAIL ROLE BADGE
Swarm Supervisory Audit Log Real-time AMCP Feed
🎮 STEAM FREE-TO-PLAY EDITION • APPID: 3133700 • 100% FREE / ZERO PAY-TO-WIN

HiveMind: Exoplanet Hunter & Cyber-Defense Hack-Miner

Play for free on Steam! Uncover habitable exoplanets from Kepler & JWST deep-space surveys, raid multi-expert MoE bosses, and run the Zero-Intrusion BOINC Daemon to share idle RTX GPU compute for science & planetary intelligence.

Steam Store: store.steampowered.com/app/3133700 Depot: 3133701 NON-MONETARY COMPUTE RECIPROCITY: EARN HMT COMPUTE POINTS
memory

RTX Zero-Intrusion BOINC Daemon

RUNNING (IDLE INFERENCE)

Shares GPU compute exclusively when you are not gaming. Auto-throttles when GPU utilization < 25% or system is idle. When a game launches, yields 100% of VRAM in <1.5s so your FPS never drops.

DETECTED GPU
NVIDIA RTX 2070 (8GB)
IDLE TRIGGER
< 25% LOAD / 30s AFK
GAME PREEMPTION
< 1.5s Target SLA
THERMAL CAP
54.2°C (Max 68°C)
VRAM OVERHEAD
1.28 GB / 8.00 GB
VERIFIED WORKUNITS
142 COMPLETED
military_tech

Compute Points Retribution Ledger

EPOCH #412
TOTAL RETRIBUTED POINTS
35,500 HMT
Beneficiary: callensxavier@gmail.com
PENDING / UNCLAIMED
+450 HMT
3 Verified Chunks
Synchronized Steam Achievements & Badges Steamworks v1.0.0
🪐
Exoplanet Transit Hunter
Fold 5 Kepler BLS Light Curves
🚀
RTX 100 Inferences
Serve 100 MoE Speculative Tokens
🐝
Golden Forager Bee
Earn 10,000 HMT Compute Points
🛡️
Byzantine Firewall Guard
Block 10 Malicious Swarm Payloads

calculate Interactive Gamer RTX Yield & Science Reward Calculator

Calculate how many HMT points, MoE tokens, and Kepler exoplanet transits your RTX GPU will process while you sleep or study:

OPTIMIZED FOR CONSUMER RTX
8 HOURS / DAY
1h (Casual) 8h (Overnight) 24h (Dedicated)
ESTIMATED DAILY YIELD
2,400 HMT / day
• MoE Inference: 38,400 tokens/day
• Kepler Transits: 48 light curves
• Low-Power Load: < 48W Average TDP

⚡ 3-STEP INSTANT SETUP FOR RTX GAMERS

1. Download Free Steam Tool

Install directly via Steam (AppID 3133700) or download the portable .zip package above.

2. Run Background Daemon

Launch Launch_HiveMind_SteamTool_RTX.bat. It auto-detects CUDA & Tensor Cores.

3. Play & Earn Reciprocal Points

Game normally! The daemon auto-pauses in <1.5s when games start and resumes when you are AFK.

WEBGPU TSL COMPUTE + 3D GALAXY ENGINE ★ Powered by dgreenheck/webgpu-galaxy ↗ SECTOR: KEPLER-186F HABITABLE ZONE

🌌 Interactive WebGPU Spiral Galaxy & Exoplanet Swarm Explorer

Real-time logarithmic spiral arm simulation leveraging dgreenheck/webgpu-galaxy compute physics: deterministic hash(seed) PRNG, Keplerian differential rotation (1.0 / (r * 0.1 + 1.0)), 3D raycast mouse repulsion (applyMouseForce), Hooke's law spring restoration (applySpringForce), and volumetric dust clouds.

SIMULATION FPS
60 FPS
STAR UNIFORMS
750,000
ASTRO BOUNTY
4,900 HMT
Kepler-186f Habitable Zone (Cygnus Sector)
Bee Expert: Astro-Exoplanet-SETI-Bee-v2.6 (Cell #3)
Transit Period: 129.9 d • SNR: 14.8σ • Cursor 3D: (0.0, 0.0, 0.0)
🖱️ Drag to Orbit 3D Camera • Scroll to Zoom • Hover to Repel Stars & Dust Clouds (Hooke's Spring k=2.0)
🪐 Warp to Sector:
✓ WEBGPU GALAXY SIMULATION READY — dgreenheck/webgpu-galaxy TSL Kernels Loaded
Select any galactic exoplanet sector above or tune the spiral arm parameters on the right, then click SCAN SECTOR & CLAIM HMT BOUNTY to fold transit lightcurves with La Ruche Cell #3.
public USS HIVEMIND NCC-31337 v4.6 • 10 STITCH EXOPLANET HABITABILITY & LIFE DOSSIER
CHAPTER 1 INTRO ONCE • ASTRO-BEE MoE
« Chapitre 1 : Passerelle. Ciblez une étoile ou sélectionnez l'une des 10 exoplanètes ci-dessus. Analyse Kepler-186f : monde rocheux froid (1.17 R⊕, 188 K, orbite habitable 130 j). Vie possible : plantes sombres captant l'infrarouge et bactéries géothermiques près des volcans. »
Kepler-186f Analysis: Cold outer-HZ rocky world (1.17 R⊕, 188 K, 130-day orbit). How life could live: dark infrared-absorbing flora and geothermal microbes near volcanic vents.
dgreenheck/webgpu-galaxy

GalaxySimulation Uniform Controls

1. Galaxy Structure (Spiral Arms)
Spiral Arm Count (armCount) 2
Spiral Tightness (spiralTightness) 1.75
Kepler Rotation Speed (rotationSpeed) 0.10
Arm Width 2.25
Randomness 1.80
Galaxy Radius 13.0
Thickness 3.0
2. Star Density Colors & Dust Clouds
Star Count Uniform (starCount) 750,000
3. Mouse Repulsion & Hooke's Spring
Mouse Force 7.0
Mouse Radius 10.0
BROWSER IDE STUDIO v2.9 • LOCAL & CLOUD RUN SYNC 🚀 Open USS HIVEMIND NCC-31337 v4.6 Full Bridge ↗

💻 HiveMind v2 Browser IDE & Live Bee-MoE / WebGPU Shader Studio

Edit, validate, and execute Astro-Bee Top-2 Sparse MoE routers, WebGPU WGSL shaders, and Stitch v2 UI components right in your browser, or open them directly in Google Antigravity IDE, VS Code, Cursor, or Windsurf.

SANDBOX RUNS
12
DEV REWARDS
2,400 HMT
✓ BROWSER IDE KERNEL READY — Zero-Trust Cyber WAF & Astro-Bee Sandbox Armed
Click RUN IN BROWSER IDE SANDBOX above to execute the active buffer and inspect the exoplanet habitability output.
Live Split-Pane Inspector

Widget Render & USS HIVEMIND NCC-31337 v4.6 Bridge

HOT-RELOAD ACTIVE
⚡ LIVE BROWSER IDE WIDGET PREVIEW
Target: K2-18 b (Hycean World • 279 K)
JWST NIRSpec: CH4 + CO2 Confirmed (98.4% SNR)
🛸 USS HIVEMIND NCC-31337 v4.6 — 5-STAGE TACTICAL COCKPIT (/cockpit-v46)
Full ↗