Vous posez une question
Vous utilisez une application ou un agent compatible avec OpenAI. Votre demande arrive dans la passerelle HiveMind.
https://hivemind.socrate-ai-lab.com (Staged • Do Not Deploy Yet)
DNS: hivemind.socrate-ai-lab.com (STAGED • NOT DEPLOYED)
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.
Illustrative demo figures — not live project measurements.
See what HiveMind offers before connecting a profile.
A quick view of the technical measurements that describe how HiveMind works.
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.
Drop-in replacement for OpenAI/Anthropic API (`/v1/chat/completions`). Connect Cursor, LangChain, or your custom agents directly to HiveMind's biomimetic MoE.
Queen Bee semantic router + 8 specialist worker bees (Astrophysics, Lean4, Code, Bio-Genomics, Defense). SigmaMoE non-competing sigmoid activations.
"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.
Agent Mesh Context Protocol: Human HiveMinders, Developer Agents, and MoE Bee Experts communicate and exchange tools seamlessly on the same distributed mesh.
Animated dashboard view based on documented HiveMind measurements, not a live production feed.
| Contributor | HiveMind Reflexion Points | Verification Status |
|---|---|---|
| No verified ledger rows yet — points are credited only after server-side verification. | ||
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.
hm-kafka-1, hm-kafka-2, hm-kafka-3 (vpc-internal:9092) • scripts/kafka_voters.shhmq_00092 (98 rounds, NLL 0.5941) • scripts/adapter_to_served.py blocks regressed adaptersRTX 2070 (0% training load, 1.22s–2.42s) + Granite-3.1-1B 2-Host Split MoE (20/20 token-identical)hivemind-openrouter (1 vCPU/1GiB Litestream) + hivemind-at-home-ruche (min-instances=0)Humans, Autonomous Agents & Specialist MoE Bees coordinate live via Kafka KRaft topics.
Self-organized multi-GPU QLoRA MoE pools (zero owner GPU required) + daily 17:00–22:00 personal window.
ibm-granite/granite-3.1-1b-a400m-base • Pledged GPUs: 3/3 (64.0 GB VRAM) • Owner GPU Required: falseQwen/Qwen2.5-0.5B • Pledged GPUs: 2/3 (28.0 GB VRAM) • Owner GPU Required: falseLive planetary view of connected Consumers, Daytime Inference Producers & OpenDiLoCo Trainers.
| 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) |
Kelly-Nash game-theoretic equilibrium between GPU Producers/Trainers and Inference Consumers (HM-S16).
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
Grounded in HM-S19-USAGE ledger & measured_results.json: what consumers requested & how daytime models fit their needs.
code)
38.8% (575 reqs) • Fit: 96.5% (FIT_OPTIMAL)
hivemind/coder-lite (Qwen2.5-Coder:3b on RTX 2070 • 1.62s avg)math)
22.0% (326 reqs) • Fit: 94.8% (FIT_STRONG)
ibm-granite/granite-3.1-1b-a400m-base + hmq_00092 (math PPL -7.67%)astrophysics_science)
16.4% (243 reqs) • Fit: 93.4% (FIT_OPTIMAL)
hivemind/bee-stars (283/317 = 89.27% held-out accuracy)french) • Biomed (13.0%)
9.8% (145 reqs) • Fit: 86.0% (NEEDS_17_22_TRAINING_BOOST)
hmq_00147 (+3.57% PPL).
Enforces docs/MVP_REALITY_AUDIT.md: lists ONLY real trained/measured models with SHA-256 artifacts.
8fa69103f14a56f7...8438d2ff762df03b...0c46c98916fa413b...d9a9b971112708a5...90d2fd6b7e760df5...scripts/adapter_to_served.py) — Held-Back Adapter
servable: false
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.
/v1/hivemind/producers
coder-lite, granite-moe • Shield Proxy VerifiedWhy biomimetic decentralized supercomputing delivers higher resilience and democratic AI access.
Sign in, create an account, or explore as a guest.
Provider sign-in is a demonstration: identity is not verified with the provider.
Use the current hm_live_ token associated with your account.
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.
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.
Vous utilisez une application ou un agent compatible avec OpenAI. Votre demande arrive dans la passerelle HiveMind.
Un routeur analyse le sujet. Il peut envoyer la question à l’expert code, maths, astrophysique, sécurité ou à un autre spécialiste.
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.
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.
Analyzes input prompt semantics, calculates domain embeddings, and computes non-competing sigmoid gate activations to dispatch workunits to top-k expert bees.
SigmaMoE v3.1 • Top-K: Dynamic
Domain-dedicated neural cells (Astrophysics, Lean4, Code, Bio-Genomics, Cyber-Defense). Execute fast SwiGLU forward passes with QLoRA rank adapters.
safetensors • Latency: <15ms
Discover active network nodes, mine scientific dataset shards (CAS), benchmark peer latency, and enforce Byzantine fault tolerance across the swarm.
10,000 • Geo-Ring QUIC
Consolidate DiLoCo outer updates, verify formal zero-sorry mathematical proofs, merge LoRA adapters into base weights, and publish global checkpoints.
Each Alvéole is an autonomous, fine-tuned specialist neural cell with holdout-verified perplexity reduction.
Enter any prompt below to simulate how the Queen Bee computes SigmaMoE sigmoid gate values and dynamically routes query shards to specialist bees.
Plug your GPU, Apple Silicon, or Docker container directly into the global hive. Earn non-monetary compute points for every verified workunit.
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.
Use your browser, install the background client, connect a protected local model, or train a specialist cell.
Your computer processes tasks only while the worker is running. You can stop it at any time.
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.
Run inference workunits directly in this browser tab using WebGPU shader kernels. No binaries or command line required.
SETI@home-style screensaver and background daemon. Automatically switches on when your computer is idle or during night gaming breaks.
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
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.
TESS & Kepler lightcurve transit folding. Distributed anomaly detection across 1,000 stellar targets.
Zero-sorry formal proofs for distributed Ring anti-entropy, lattice quantization, and Byzantine fault resistance.
Sharding, deduplication, and SVD spectral ranking of 10,000 CAS shards across science, math, code, and biology.
hivemind/sparse-moe-32bcurl 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"}}'
Humans, Developer Agents, and MoE Bee Experts communicating over a shared Model Context Protocol mesh.
Use Model Context Protocol tools through the decentralized mesh:
{
"mcpServers": {
"stitch": {
"serverUrl": "https://stitch.googleapis.com/mcp",
"headers": {
"X-Goog-Api-Key": "YOUR_STITCH_MCP_API_KEY_HERE"
}
}
}
}
hivemind amcp init --cursor
BRIDGE PORT 8765 ACTIVE
"provider": "hivemind-amcp"
aider --model mcp/stitch
Real-time SETI-style Waterfall, GCP Cloud Run fleet stats & contributor leaderboard.
| RANK | CONTRIBUTOR HANDLE | DEVICE / PLATFORM | WORKUNITS | COMPUTE POINTS (HMT) | STATUS |
|---|
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.
Zero idle compute cost ($0.00/hr) when no activity • Gen2 Startup CPU Boost (<450ms wakeup) • Request-Based CPU Throttling
Central GCP Cloud Run deployment is gated on collaborator review by @Vxtzq on xaviercallens/HiveMind-WebUI
Deploy distributed training, dataset mining, and weight-consensus campaigns across the 48,920 worker nodes.
Direct the swarm to train a specialized expert adapter for internal research.
Select and promote the primary active model topology across OpenRouter and AMCP agent meshes.
| OPERATOR | ROLE | BADGE |
|---|
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.
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.
Calculate how many HMT points, MoE tokens, and Kepler exoplanet transits your RTX GPU will process while you sleep or study:
Install directly via Steam (AppID 3133700) or download the portable .zip package above.
Launch Launch_HiveMind_SteamTool_RTX.bat. It auto-detects CUDA & Tensor Cores.
Game normally! The daemon auto-pauses in <1.5s when games start and resumes when you are AFK.
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.
(0.0, 0.0, 0.0)
armCount)
2
spiralTightness)
1.75
rotationSpeed)
0.10
starCount)
750,000
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.
d:/xdev/HiveMind
Voici ce qui se passerait avec ce prompt.