128 GPUs Online

Real Compute. Real Data.
Real Outcomes.

We own 128 Blackwell GPUs. We curate 1.2 million verified intelligence pairs. Every pair passes 8 gates — no overrides, no exceptions, no shortcuts. Sovereign infrastructure for vertical AI that works in the field, not just on a benchmark.

Explore Capital Markets → See How We Build
128
GPUs Owned
1.2M
Verified Pairs
$8B
In Closed Deals
0%
Override Rate
Checking API...
What We Do

Curate. Build. Deploy.

We produce verified intelligence on hardware we own. No rented clouds. No third-party data. No middlemen between your model and the truth.

👑

The Refinery

Raw signal in. Verified intelligence out. 8 gates. Zero overrides. If it doesn't hit platinum, it doesn't ship. 1.2 million pairs and counting.

1.2M pairs
85+ specialties
0% override
See the Pipeline →
🏛

Capital Markets

The CRE all-in-one. 643K verified intelligence pairs, 19 live skills, AI analyst, edge deployment. One vertical — underwriting to close — fully built.

643K CRE pairs
19 live skills
9 asset types
Explore Capital Markets →
🚀

Your Vertical

You bring the domain. We bring 128 GPUs and an 8-step pipeline that has never been overridden. CRE, medical, pharma, aviation — or whatever you need built.

128 GPUs owned
Any vertical
12TB VRAM
Request a Build →
Data Quality

SwarmCurator

Every pair in the Gold Vault passes an 8-step curation pipeline. No shortcuts. No overrides. The final verdict is unappeallable — if it doesn't hit Platinum, it doesn't ship.

1 🔌 Source Raw generation via Together.ai cooks & domain grinders
2 ⚙ Grind Specialty tagging, domain formatting, task typing
3 📋 Schema Structural validation, JSON schema, field completeness
4 🔎 CoVe Chain of Verification — 5 criteria, scored 1-10
5 ⚖ Judge SwarmCurator PASS/FAIL verdict on every pair
6 🛠 Trajectory Multi-step reasoning path verified end-to-end
7 🔀 Cross-Val Dedup, consistency check, adversarial probing
8 👑 Platinum Final promotion. Unappeallable. Ships to training.
1.2M
Platinum Pairs
8
Pipeline Steps
5
CoVe Criteria
85+
Specialties
0%
Override Rate
SwarmCurator verdict pipeline: accuracy 9.2 │ completeness 9.4 structure 9.7 │ relevance 9.1 sft_quality 9.5 │ trajectory VERIFIED ────────────────────────────────────────── verdict: PLATINUM PROMOTED │ override: DENIED status: unappeallable │ ships to training

Custom Curation Builds

CRE, pharma, aviation, legal, finance — any vertical. You bring the model spec. We build the dataset. Same 8-step SwarmCurator pipeline, tuned to your domain, your taxonomy, your quality bar. Platinum pairs delivered — ready to train.

You Define
Domain, specialties, task types, volume
Tell us the vertical, the model size, and what the pairs need to do.
We Build
Generate, grind, validate, judge, promote
Full 8-step pipeline on our 128-GPU sovereign compute. No third-party data.
You Ship
Platinum pairs, eval sets, training configs
Delivered as JSONL with CoVe scores, trajectory labels, and train/eval splits.
Request a Custom Build →
Model Fleet

The Fleet

Vertical AI models trained on sovereign compute. 1.24M platinum pairs. RTX PRO 6000 Blackwell + RTX 3090 Ti. Every model earns its slot.

💊

SwarmPharma-35B

Qwen3.5-35B-A3B · bf16 LoRA
Clinical pharmacology. DDI detection, dosing, PK/PD, 5-step trajectory. 25,629 pairs. Zero quantization loss.
Sealed v1
🏛

SwarmCRE-35B

Qwen3.5-35B-A3B · bf16 LoRA
Commercial real estate. Underwriting, IC memos, deal analysis. 643K CRE pairs. v2 training in progress.
Training v3
🩺

SwarmCurator-27B

Qwen3.5-27B · Q8_0 · Blackwell
Quality gate + curation. Absorbs judge role. CRE, medical, aviation pair generation. 96GB VRAM.
Live on GPU 1
🔎

SwarmCurator-9B

Qwen3.5-9B · F16 · 3090 Ti
Ops specialist. 46K pairs (methodology + pipeline ops). Research, safety, signal generation.
Live on GPU 0
📚

SwarmResearch-32B

Qwen2.5-32B · bf16 LoRA
Safety & research. 35,499 pairs. Multi-domain analysis, aviation, core verticals.
Trained v1
👑

SwarmCRE-9B "Morey"

Qwen3.5-9B · Chat & Voice
The CRE flagship. Underwriting, IC memos, deal analysis. 28 SwarmSkills. Live API.
Live
🔭

BeeMini-3B

Qwen2.5-3B · Q4_K_M · 1.8GB
Skill router. 98.3% valid JSON, 79.7% domain accuracy. 60K pairs. Runs on anything.
Live
📡

SwarmSignal-2B

Qwen3.5-2B · Q4_K_M · Jetson
Edge signal classifier. P1-P5 priority scoring. 15-min cycles. Jetson Orin Nano 8GB.
Live on Edge
Agent Economy

28 SwarmSkills

Each skill is schema-validated, quality-gated by SwarmCurator, and callable via a single API endpoint. Every output becomes training data. Two verticals live.

Commercial Real Estate · 19 Skills
💼broker_senior
👤broker_junior
🔍intelligence_query
🎲bookmaker
📋deal_tracker
💻developer
📡signal_scraper
💰investor
🔃exchange_1031
📊market_report
🎯lead_scorer
✉email_composer
📈comp_analyzer
📄rent_roll_analyzer
🏦debt_analyzer
💵tax_assessor
📍site_selector
🚀portfolio_optimizer
📰news_digest
Pharma & Clinical · 9 Skills
💊drug_interaction
📋med_reconciliation
🔢dose_calculator
🩸anticoag_advisor
⚠allergy_check
🧬pgx_advisor
🚨adverse_event
👶pregnancy_safety
💉med_safety
Edge Compute

BeeBox

A sovereign AI appliance on your desk. Local inference, plug-and-play skills, and fleet escalation when you need it.

┌─────────────────────────────────────────────┐ │ BeeBox · Sovereign Edge AI │ ├─────────────────────────────────────────────┤ │ │ │ BeeMini-3B → Routes every request │ │ Judge-4B → Quality gate (local) │ │ Skill Store → Plug-and-play skills │ │ │ │ 72% local · 28% fleet escalation │ │ │ │ ─── fleet call ──→ SwarmHQ (9B+) │ │ ←── response ──── GPU Cluster │ │ │ └─────────────────────────────────────────────┘
$499
Target Price
2-4B
Local Models
16 TOPS
NPU Accel
Infrastructure

The Stack

From raw data to edge inference — every layer is purpose-built.

EDGE BeeBox · Jetson · Mac Mini · Zima BeeMini-3B router → Judge-4B gate Q4_K_M GGUF · 2-6GB VRAM · USB4/10GbE API router.swarmandbee.com POST /skill/{name} → route → execute → judge → store Cloudflare Workers · D1 · R2 · Vectorize COMPUTE 128× RTX PRO 6000 Blackwell · 96GB each · 12,288GB VRAM We own the racks. We cook the rails. bf16 LoRA · Unsloth · llama.cpp · vLLM · fp8 inference FLEET SwarmPharma-35B · SwarmCRE-35B · Curator-27B · Curator-9B · BeeMini-3B · Signal-2B SwarmCurator 8-step pipeline → Platinum pairs only Mamba-Transformer · MoE · Q4_K_M GGUF · sovereign DATA Gold Vault · 1.2M pairs · 85+ specialties R2 buckets: sb-cre · sb-medical · sb-aviation · sb-core · sb-drone CoVe verification · 6 deterministic gates · trajectory labels TRAIN Factory · Together.ai cooks · Blackwell training Phase pipeline: generate → judge → promote → train → eval Qwen3.5 base · LoRA r=64 · packing · multi-phase
Data Infrastructure

From Compute to Delivery

Four layers. Zero middlemen. Raw GPU cycles become verified intelligence objects delivered however you need them.

128
Blackwell GPUs
1.2M
Verified Pairs
250K
Pairs / Month
12TB
Total VRAM
1
Real Compute Layer
128 RTX PRO 6000 Blackwell GPUs vLLM inference fleet Swarm data factory 96GB per GPU
↓
2
Data Curation Layer
1.2M verified intelligence pairs 8-gate verification pipeline Trajectory mutation engine Entropy scoring
↓
3
Data Delivery Layer
API access Dataset downloads Vector embeddings Model training packs Live signal streams
↓
4
Vertical AI
Commercial Real Estate Medical Pharma Aviation
🔌
API Access
REST endpoints, real-time queries, skill calls
📦
Datasets
Platinum JSONL, train/eval splits, CoVe scored
🔭
Vectors
BGE-Base embeddings, Cloudflare Vectorize
⚡
Training Packs
LoRA configs, model cards, eval suites
📡
Signal Streams
Live P1-P5 market intel, 15-min edge cycles
Powered By

Built With

Local sovereignty. Open-source foundation. Every tool in the stack is chosen because we own the signal — not rent it.

🤖 Qwen3.5 BASE
Alibaba's Qwen3.5 family — Mamba-Transformer hybrids and MoE architectures. The base brain for every SwarmCRE model.
Qwen3.5-122B · Founder FTW, fp8, 2 GPUs
Qwen3.5-9B · Mamba-Transformer flagship
Qwen3.5-4B · Compact edge
Qwen2.5-3B · BeeMini router
Qwen3.5-0.8B · Nano, CPU inference
Qwen on HuggingFace →
⚡ Unsloth TRAIN
1.5x faster fine-tuning, 50% less VRAM. Every LoRA adapter in the fleet is trained with Unsloth on Blackwell GPUs.
bf16 LoRA r=64 · full precision
packing=True · 6x throughput
FA2 · Flash Attention 2
5GB VRAM · min for Qwen3.5-2B LoRA
GitHub →
🐦 llama.cpp SERVE
GGUF quantization and inference for edge deployment. Every model ships as Q4_K_M for BeeBox, Jetson, and desktop.
Q4_K_M · primary quant format
llama-server · GPU inference
llama-quantize · model export
sm_86 / sm_120 · Ampere + Blackwell
GitHub →
☁ Cloudflare EDGE
Workers, D1, R2, Vectorize, Pages. The entire API layer and data plane runs on Cloudflare's edge network.
Workers · SwarmSkills runtime
D1 · events, entities, memory
R2 · Gold Vault storage
Vectorize · BGE-Base embeddings
Cloudflare Docs →
🚀 Together.ai COOK
Serverless inference API for data cooking. Generates and judges millions of pairs through the SwarmCurator pipeline.
Qwen3-Next-80B-A3B · generation
Qwen3-235B-A22B · quality rewrite
Llama-4-Maverick · judge verdicts
15 workers · per cook run
together.ai →
💻 NVIDIA GPU
128 RTX PRO 6000 Blackwell GPUs. 96GB each. 12,288GB total VRAM. We own the racks. We cook the rails.
RTX PRO 6000 · 96GB, sm_120
RTX 3090 Ti · 24GB, sm_86
Jetson Orin Nano · 8GB, edge
CUDA 12.8+ · Blackwell support
nvidia.com →
⚡ Supabase DB
Postgres backend for training runs, model registry, and operational state. Auth, realtime, and storage in one platform.
training_runs · job tracking & status
model_registry · artifacts & versions
Auth · API key management
Realtime · live training updates
supabase.com →
Who We Are

$8 Billion in Closed Deals.
Then We Built the Machine.

Donovan Mackey spent 30 years as Senior Managing Director at Marcus & Millichap — leading the national industrial team, closing $8B in real transactions across cold storage, supply chain logistics, and economic development.

Every model, every skill, every gate in the Swarm pipeline maps to a real step in how deals actually close. This isn't AI built by engineers guessing at industry. This is the deal machine — automated. Donovan Mackey Founder & CEO — Swarm & Bee
By the Numbers

The Signal Is Real

643K
CRE Platinum Pairs
"Math-verified. Every underwriting calc checked."
432K
Medical Pairs
"85 specialties. Trajectory-enhanced. DDI-verified."
45K
Aviation Pairs
"157 specialties. Operational relevance scored."
0%
Override Rate
"Six deterministic gates. No LLM judge. Code decides."
Questions

FAQ

What is Swarm & Bee?
Swarm & Bee builds vertical AI models and curates platinum-grade training data. We own 128 RTX PRO 6000 Blackwell GPUs (12TB VRAM) and run the full stack — from data generation to model training to edge deployment. Our first vertical is commercial real estate (SwarmCRE), with medical, aviation, and custom verticals available.
What is SwarmCRE?
SwarmCRE is our commercial real estate AI franchise — a family of models from 0.8B (edge/IoT) to 122B (Founder FTW). The flagship is SwarmCRE-9B "Morey," a Mamba-Transformer hybrid trained on 643K platinum CRE pairs. It handles underwriting, IC memos, deal analysis, debt sizing, 1031 exchanges, market intelligence, and more via chat or voice.
What makes your training data different?
Every training pair passes our 8-step SwarmCurator pipeline: Source, Grind, Schema, CoVe (Chain of Verification — 5 criteria scored 1-10), Judge (PASS/FAIL verdict), Trajectory (multi-step reasoning), Cross-Validation (dedup + adversarial probing), and Platinum promotion. The final verdict is unappeallable — 0% override rate across 1.2M pairs.
Can you build custom training data for my vertical?
Yes. We offer Custom Curation Builds for any domain — CRE, pharma, aviation, legal, finance, or something new. You define the specialties, task types, and volume. We run the full 8-step pipeline on 128 GPUs. You receive platinum JSONL with CoVe scores, trajectory labels, and train/eval splits. Contact us to scope a build.
What is BeeBox?
BeeBox is our sovereign edge AI appliance — a plug-and-play device that runs BeeMini-3B (skill router) and SwarmCRE-4B (quality gate) locally. It handles 72% of requests on-device and escalates 28% to the SwarmHQ GPU fleet. Target price: $499. Designed for brokerages, investors, and teams who want CRE AI without cloud dependency.
How do SwarmSkills work?
SwarmSkills are schema-validated AI agent skills callable via a single API endpoint: POST /skill/{name} at router.swarmandbee.com. There are 28 live skills across CRE and pharma verticals — from broker analysis to drug interaction screening. Every skill output is quality-gated by SwarmCurator and becomes training data for the next model generation.
Get In Touch

Build With Us

Interested in SwarmCRE, the API, BeeBox, or joining the fleet? Drop a line.