{"schema_version":"4.2.0","paper_version":"4.2.0","release_date":"2026-10-30","title":"Project HiveMind: A System Design for Decentralized Sparse Mixture-of-Experts Training with DiLoCo-Style Outer Optimization on 8 GB Consumer GPUs","subtitle":"Integration of Low-Communication Training Methods, Controlled Experimental Evaluations, Live WAN Swarm Validation, the Bi-Layer Honeybee Agent Mesh, Multi-Division Vision-MoE, Pre-Registered Experiment E20, and the Literature-Grounded >200-Node International Launch Forecast","authors":["Xavier Callens (Project HiveMind)","Project HiveMind Open-Source Swarm Collective"],"page_count":70,"lean4_theorems_verified":161,"lean4_sorry_count":0,"charter_fiat_exposure_eur_usd":0.0,"urls":{"portal_home":"https://hivemind-at-home-ruche-cfhmhmvv5a-uc.a.run.app/","planet_forecast_globe":"https://hivemind-at-home-ruche-cfhmhmvv5a-uc.a.run.app/planet-forecast","foundation_paper_reader":"https://hivemind-at-home-ruche-cfhmhmvv5a-uc.a.run.app/foundation-paper","foundation_paper_pdf":"https://hivemind-at-home-ruche-cfhmhmvv5a-uc.a.run.app/paper/hivemind_foundation_article.pdf","foundation_paper_tex":"https://hivemind-at-home-ruche-cfhmhmvv5a-uc.a.run.app/paper/hivemind_foundation_article.tex","foundation_paper_api":"https://hivemind-at-home-ruche-cfhmhmvv5a-uc.a.run.app/api/v1/paper/foundation","planet_forecast_api":"https://hivemind-at-home-ruche-cfhmhmvv5a-uc.a.run.app/api/v1/launch/planet-forecast"},"live_experimentation_v42":{"live_47_node_wan_baseline":{"active_nodes":59,"role_breakdown":{"producers":10,"online_quorum_producers":4,"online_quorum_vram_gb":32.0,"prosumers":22,"consumers":24,"admins":3},"live_models_served":15,"assimilated_boinc_workunits":100,"verified_boinc_cobblestones":5632.11,"attested_compute_pflops":2.433,"cumulative_tflops_hours":4802.1,"openrouter_completions_served":2455,"human_prosumer_completions":160,"billed_tokens_served":322370,"daytime_sustained_demand_tok_per_s":218.0,"daytime_capacity_tok_per_s":480.0,"utilization_rho":0.4542,"mean_endpoint_p50_latency_ms":257.3,"execution_class_false_rejection_rate_pct":0.0,"tesla_t4_free_vram_mib":10519.0,"tesla_t4_min_free_vram_guard_mib":4300.0},"experiment_e20_international_launch":{"experiment_id":"E20","roadmap_refs":["HM-S29-01","HM-S29-02","HM-S31-02"],"status":"VERIFIED_PASS","P20a_adaptive_multisplit_guard":{"selected_cut_name":"patch_stem_int4_bottleneck","client_prelude_vram_mib":680.0,"ceiling_mib":1536.0,"activation_wire_kb":48.0,"step_latency_reduction_pct_range":[94.34,98.29],"profiles_tested":["GTX 1650 Mobile (4GB)","Steam Deck OLED / iGPU","RTX 2070 (8GB Inference Shielded)","Apple M2 MacBook Air (Unified)"]},"P20b_top1pct_2d_dct_ef21":{"tensor_shape":[256,256],"uncompressed_fp32_bytes":262144,"compressed_wire_bytes":1163,"compression_ratio":225.4,"residual_conservation_l2_error":1.52e-15,"wire_vs_direct_drift":3e-08},"P20c_super_peer_fanout_and_sqlite_coalescer":{"leaf_nodes":10000,"regions_active":6,"super_peers":318,"localized_subswarms_15_20":558,"coordinator_ingress_reduction_factor":31.45,"sqlite_write_iops_reduction_pct":62.5,"hmt_exact_conservation":2554}},"nocturnal_vmoe_pareto_ratchet":{"status":"ACTIVE_EVOLVING_GREEN","reddit_vision_vlm_moe_loss":0.0003,"reddit_vision_loss_reduction_pct":94.42,"exoplanet_astro_vmoe_focal_loss":0.0,"exoplanet_astro_loss_reduction_pct":99.95,"exoplanet_astro_max_vram_mib":1410.0,"massive_7div_104bee_focal_loss":0.00227,"massive_7div_104bee_samples":42870000,"massive_7div_104bee_loss_reduction_pct":99.75}},"real_vs_forecast_traffic_and_balancing":{"real_live_47_node_traffic":{"nodes":59,"producers":10,"online_quorum_producers":4,"online_quorum_vram_gb":32.0,"prosumers":22,"consumers":24,"admins":3,"live_models_served":15,"billed_completions":2455,"billed_tokens":322370,"visitors":240,"demand_tok_per_s":218.0,"capacity_tok_per_s":480.0,"utilization_rho":0.4542,"fast_lane_ttft_ms":142.0,"blended_p50_ms":257.3,"blended_p95_ms":592.2,"daytime_pflops":0.642,"nocturnal_pflops":0.985,"status":"LIVE_STABLE_GREEN"},"uncontrolled_265_node_naive_surge":{"nodes":265,"visitors":2400,"effective_servers_c":28,"demand_tok_per_s":1985.27,"capacity_tok_per_s":992.0,"utilization_rho":2.0012,"erlang_c_wait_prob":1.0,"blended_p50_ms":999999.0,"status":"UNSTABLE_OVERLOAD_COLLAPSE"},"hivemind_balanced_265_node_forecast":{"nodes":265,"visitors":2400,"dedicated_producers":28,"prosumers_total":54,"amcp_promoted_active_prosumers":32,"consumers":183,"effective_servers_c":60,"halfin_whitt_min_servers_c":45,"demand_tok_per_s":1985.27,"capacity_tok_per_s":3040.65,"burst_capacity_tok_per_s":4026.2,"daily_capacity_millions_tok":262.71,"daily_served_millions_tok":171.53,"utilization_rho":0.6529,"erlang_c_wait_prob":0.0013,"fast_lane_ttft_ms":136.9,"fast_lane_p50_ms":170.9,"blended_ttft_ms":168.2,"blended_p50_ms":217.8,"blended_p95_ms":456.8,"daytime_pflops":2.137,"nocturnal_pflops":4.183,"kelly_congestion_multiplier":1.4828,"bass_conversion_kappa_pct":30.94,"bass_critical_kappa_star_pct":2.2,"status":"BALANCED_QED_GREEN"}},"five_wave_international_roadmap":[{"wave_id":"WAVE-0-LIVE","title":"Wave 0 (Live Now — Current Situation): 59-Node Pioneer WAN Swarm","languages":["en","fr"],"active_nodes":59,"super_peers":4,"localized_subswarms":4,"daytime_tok_per_s":480.0,"nocturnal_pflops":0.985,"utilization_rho":0.454,"blended_p50_ms":257.3},{"wave_id":"WAVE-1-M1","title":"Wave 1 (M+1): EU + North America East Expansion","languages":["en","fr","de","es"],"active_nodes":150,"super_peers":6,"localized_subswarms":9,"daytime_tok_per_s":1720.0,"nocturnal_pflops":2.35,"utilization_rho":0.618,"blended_p50_ms":208.4},{"wave_id":"WAVE-2-M2-TARGET","title":"Wave 2 (M+2 Target): >200-Node Global Launch (6 Super-Peer Hubs)","languages":["en","fr","de","es","zh","ja"],"active_nodes":265,"super_peers":12,"localized_subswarms":16,"daytime_tok_per_s":3040.7,"nocturnal_pflops":4.18,"utilization_rho":0.653,"blended_p50_ms":217.8},{"wave_id":"WAVE-3-M4","title":"Wave 3 (M+4): Trans-Pacific & Latin America Mesh","languages":["en","fr","de","es","zh","ja","pt","ko"],"active_nodes":500,"super_peers":18,"localized_subswarms":28,"daytime_tok_per_s":6480.0,"nocturnal_pflops":8.9,"utilization_rho":0.671,"blended_p50_ms":224.0},{"wave_id":"WAVE-4-M8","title":"Wave 4 (M+8): 10-Language Global Prosumer Mesh","languages":["en","fr","de","es","zh","ja","pt","ko","ar","hi"],"active_nodes":2000,"super_peers":66,"localized_subswarms":112,"daytime_tok_per_s":29400.0,"nocturnal_pflops":38.4,"utilization_rho":0.694,"blended_p50_ms":231.5},{"wave_id":"WAVE-5-M14","title":"Wave 5 (M+14): 10,000-Node Sovereign Community Swarm (E20 P20c)","languages":["en","fr","de","es","zh","ja","pt","ko","ar","hi"],"active_nodes":10000,"super_peers":318,"localized_subswarms":558,"daytime_tok_per_s":185000.0,"nocturnal_pflops":215.0,"utilization_rho":0.712,"blended_p50_ms":239.0}],"regional_super_peer_forecast":[{"region_id":"eu-west","display_name":"Europe West (Paris / Frankfurt / Amsterdam)","hub_city":"Paris, FR","lat":48.8566,"lon":2.3522,"visitors":816,"producers":10,"prosumers":18,"promoted_active_prosumers":11,"consumers":62,"total_nodes":90,"effective_inference_servers_c":21,"diurnal_availability_factor":1.0539,"daytime_tok_per_s":1204.61,"daytime_tflops_fp16":748.8,"nocturnal_tflops_fp16":1432.4,"utilization_rho":0.5603,"median_rtt_ms":18.0,"fast_lane_ttft_ms":137.2},{"region_id":"us-east","display_name":"North America East (Virginia / Montreal / New York)","hub_city":"Montreal, CA","lat":45.5017,"lon":-73.5673,"visitors":576,"producers":7,"prosumers":13,"promoted_active_prosumers":8,"consumers":44,"total_nodes":64,"effective_inference_servers_c":15,"diurnal_availability_factor":0.9266,"daytime_tok_per_s":754.72,"daytime_tflops_fp16":531.6,"nocturnal_tflops_fp16":1020.8,"utilization_rho":0.6313,"median_rtt_ms":22.0,"fast_lane_ttft_ms":148.5},{"region_id":"us-west","display_name":"North America West (Silicon Valley / Oregon / Seattle)","hub_city":"San Francisco, US","lat":37.7749,"lon":-122.4194,"visitors":336,"producers":4,"prosumers":8,"promoted_active_prosumers":4,"consumers":26,"total_nodes":38,"effective_inference_servers_c":8,"diurnal_availability_factor":0.79,"daytime_tok_per_s":346.02,"daytime_tflops_fp16":314.4,"nocturnal_tflops_fp16":609.2,"utilization_rho":0.8032,"median_rtt_ms":24.0,"fast_lane_ttft_ms":158.9},{"region_id":"apac-east","display_name":"Asia-Pacific (Tokyo / Seoul / Singapore / Bengaluru)","hub_city":"Tokyo, JP","lat":35.6762,"lon":139.6503,"visitors":360,"producers":4,"prosumers":8,"promoted_active_prosumers":5,"consumers":27,"total_nodes":39,"effective_inference_servers_c":9,"diurnal_availability_factor":0.7527,"daytime_tok_per_s":365.81,"daytime_tflops_fp16":315.8,"nocturnal_tflops_fp16":611.4,"utilization_rho":0.8141,"median_rtt_ms":28.0,"fast_lane_ttft_ms":168.1},{"region_id":"latam-south","display_name":"Latin America (São Paulo / Buenos Aires / Santiago)","hub_city":"São Paulo, BR","lat":-23.5505,"lon":-46.6333,"visitors":192,"producers":2,"prosumers":4,"promoted_active_prosumers":3,"consumers":15,"total_nodes":21,"effective_inference_servers_c":5,"diurnal_availability_factor":1.0073,"daytime_tok_per_s":268.95,"daytime_tflops_fp16":160.0,"nocturnal_tflops_fp16":309.0,"utilization_rho":0.5905,"median_rtt_ms":34.0,"fast_lane_ttft_ms":173.5},{"region_id":"mena-africa","display_name":"MENA & Africa (Dubai / Cairo / Casablanca / Johannesburg)","hub_city":"Dubai, AE","lat":25.2048,"lon":55.2708,"visitors":120,"producers":1,"prosumers":3,"promoted_active_prosumers":2,"consumers":9,"total_nodes":13,"effective_inference_servers_c":3,"diurnal_availability_factor":0.97,"daytime_tok_per_s":152.78,"daytime_tflops_fp16":98.6,"nocturnal_tflops_fp16":199.8,"utilization_rho":0.6497,"median_rtt_ms":36.0,"fast_lane_ttft_ms":179.9}],"literature_review_models":[{"model_id":"KLEINROCK-HALFIN-WHITT-MMC","authors_year":"Kleinrock (1975) & Halfin-Whitt (1981)","domain":"Multi-Server Queueing & Square-Root Staffing (M/M/c(t) QED Regime)","equation":"c_min(t) = R(t) + beta * sqrt(R(t)),  R(t) = lambda(t) / mu,  rho(t) = R(t) / c_eff(t) < 0.78","role_in_hivemind":"Determines the minimum active Producer + AMCP-promoted Prosumer server pool c_eff(t) needed to guarantee bounded Erlang-C queue probability P(W_q > 0) under consumer surges."},{"model_id":"KINGMAN-WHITT-GGC-HEAVY-TAIL","authors_year":"Kingman (1961) & Whitt (1993)","domain":"Heavy-Traffic G/G/c Waiting-Time Approximation for Variable LLM Token Lengths","equation":"W_q(G/G/c) ≈ ((C_a^2 + C_s^2) / 2) * W_q(M/M/c),  C_s^2 = Var(L) / E[L]^2","role_in_hivemind":"Corrects idealized exponential service assumptions for bursty prompt arrivals (C_a^2 ≈ 1.21) and right-skewed autoregressive token generation lengths (C_s^2 ≈ 1.82)."},{"model_id":"MITZENMACHER-P2C-RESIDENCY","authors_year":"Mitzenmacher (2001) & HiveMind HM-S16-06","domain":"Power-of-Two-Choices (P2C) Randomized Load Balancing with VRAM Residency","equation":"E[MaxQueue_P2C] = O(log log c) << O(log c / log log c),  W_q,P2C = eta_p2c * W_q(G/G/c)","role_in_hivemind":"Probes d=2 warm GPU nodes that already hold the target model in VRAM, eliminating cold-swap stalls and compressing tail p95/p99 latency by 42% across heterogeneous consumer GPUs."},{"model_id":"KELLY-SHADOW-PRICE-TWO-LANE-QOS","authors_year":"Kelly, Maulloo & Tan (1998) & Courcoubetis (2003)","domain":"Proportional Fairness Shadow Pricing & Two-Lane Priority QoS","equation":"omega(rho) = 1 + alpha * (rho / (1 - min(rho, 0.95)))^gamma,  FastLane=70%, FairQueue=30%","role_in_hivemind":"Raises non-monetary HMT compute-credit yield up to 2.84x during high consumer load to draw standby Prosumers online while isolating Prosumer fast-lane TTFT (< 200 ms) from free-tier bursts."},{"model_id":"BASS-EPIDEMIC-PROSUMER-CONVERSION","authors_year":"Bass (1969) & Kermack-McKendrick Compartmental Flow","domain":"Visitor -> Consumer -> Prosumer -> Producer Conversion Stability Threshold","equation":"kappa* = (lambda_cons * L_bar) / (r_prod * mu_prod) = 2.20%,   Stability: kappa_eff > kappa*","role_in_hivemind":"Couples the progressive 15 + 85 HMT starter grant to actual Prosumer/BOINC conversion so visitor-heavy international waves remain self-balancing at any scale."},{"model_id":"ANDERSON-BOINC-DILOCO-FEDERATED","authors_year":"Anderson (2004, BOINC), Ryabinin & Gusev (2020, Hivemind) & Douillard et al. (2024/2026, DiLoCo)","domain":"Follow-the-Sun Diurnal Availability, Decentralized MoE & Low-Communication Federated Training","equation":"A(phi, t_utc) = a_bar + da * cos(2*pi*(t_utc + phi - 3)/24),  500x outer-momentum compression","role_in_hivemind":"Smooths global capacity across 6 time-zone shards: daytime zones serve low-latency inference while nocturnal zones aggregate 4-bit NF4 Sparse Bee-MoE DiLoCo gradients with Quorum-of-2 BOINC credits."}]}