Boundary Labs / Research Infrastructure

The Stack

Hardware specs, inference configuration, optimization history, and agent infrastructure. Everything documented. Raw numbers only.

Research Cluster

Compact two-tier research cluster with a clean split between orchestration and inference. Direct-wired, low-latency, and optimized for local-first operation under real workload conditions.

Primary Inference Node

SystemCyberPowerPC GXi3400BSTV17

CPUIntel Core Ultra 7 265F

CPU config20c/20t · 5.3GHz boost

GPU ×2RTX 5060 Ti 16GB GDDR7

ArchitectureBlackwell SM_120

VRAM total32 GB GDDR7 (TP=2)

PCIe configx8 Gen5 + x4 Gen4

RAM32 GB DDR5 (192 GB max)

Storage2 TB PCIe 4 NVMe

CUDA13.0.3

InferencevLLM + llama.cpp

GatewaySingle internal inference gateway

OSUbuntu 24.04

Orchestration Node

SystemBeelink EQI12

CPUIntel i5-1235U (12th Gen)

CPU config12c · 4.4GHz boost

RAM32 GB DDR4

Storage1 TB NVMe

OSopenSUSE Tumbleweed (since 2026-07-08)

UptimeAlways-on

Active servicesharness, Mike, Hermes, autoresearch, chronicle

Webnginx · Cloudflare tunnel

NetworkDirect-wired to inference tier

Remote accessPrivate mesh access

Backup476 GB NVMe · 3am daily · 30-day retention

Dev Workstation

SystemMacBook Air M4

RAM32 GB unified

RolePrimary dev, Claude Code CLI

Remote accessPrivate mesh access

Network — Topology

Cluster linkDirect Ethernet

ExternalPrivate mesh access

Inference accessSingle internal gateway for all clients

Web routingCloudflare tunnel → nginx

SSHPrivate-only, not exposed publicly

Inference Configuration

All inference clients route through a single internal gateway. Backends are hot-swappable without reconfiguring consuming services. The gateway exposes an OpenAI-compatible interface with model aliasing and auth.

~97

tok/s production warm (Genesis)

measured 2026-07-08 via gateway

2,436

tok/s prompt (dual GPU)

prefill throughput

32 GB

VRAM available (TP=2)

dual RTX 5060 Ti

swap used (GPU tier)

vs 4–8 GB on CPU tier

Model	Quant	Gen	Context	Server	Status	Notes
Qwen3.6-27B GPTQ INT4 + MTP n=3 (Genesis)	AutoRound INT4	~97 t/s warm (2026-07-08)	65K	vLLM (Genesis-patched)	active — production	Current default backend, restored 2026-07-03 after the Ornith trial. GPTQ-Marlin, TP=2, MTP n=3 speculative decoding, chunked prefill + prefix caching.
Ornith-1.0-35B	Q4_K_M GGUF / AEON NVFP4	~101 t/s warm / 124 peak (prod config)	131K (NVFP4)	llama.cpp / vLLM 0.23	trial ended 2026-07-03	Production trial 2026-06-26 → 2026-07-03. Lab peak 129.9 t/s short-context (2026-06-25). Won sequential tool composition vs DeepSeek V3.2 (20/20 vs 11/20, ~10× faster); DeepSeek won the suite overall by one point.
Nemotron 3 Nano 30B (MoE hybrid)	Q4_K_M	117.34 t/s med / 117.60 peak	32K	llama.cpp cuda128-clean	standby — dated peak 2026-05-20	Mamba/SSM hybrid MoE. ~3B active params/token. Requires cuda128-clean build; cuda13 crashes with MMQ CUDA error on SM_120.
AEON (Qwen3 NVFP4)	NVFP4 (ModelOpt)	~69 t/s	122K	vLLM 0.9+	stopped	Blackwell-native quantization. MTP speculative decoding n=3.
Qwen3.6-27B (SSM hybrid)	Q4_K_M	~22 t/s	65K	llama.cpp	stopped	Mamba/SSM hybrid. Fast prefill (960 t/s), slow gen (SSM bottleneck).
Gemma 4 26B (CPU baseline)	Q4_K_M	~11.7 t/s	32K	llama.cpp	historical	CPU-only orchestration-tier baseline from the pre-GPU phase. Consistent swap usage 4–8 GB.

Blackwell SM_120 fast CUDA kernel paths: q4_0 and f16 KV only. q8_0, q5_0, iq4_nl all degrade significantly. NVFP4 (ModelOpt) is Blackwell-native quantization — 4× smaller KV footprint vs bf16, requires CUDA 12.8+. This system runs CUDA 13.0.3; Qwen3-series NVFP4 (FlashInfer CUTLASS SM_120 FP4 kernels) should now be unblocked — previously blocked on CUDA 12.8. llama.cpp users: build target matters. cuda13 crashes with CUDA MMQ error on SM_120; cuda128-clean required for llama.cpp GGUF inference.

Inference Optimization — Autoresearch Log

130+ automated autoresearch runs across six model architectures using a scripted loop — a subset of the 215+-run total benchmark record. Stopping criterion: +5 tok/s improvement per iteration. Full logs at inference-research.

Key Configurations — Qwen3.6-35B-A3B MoE (dual GPU, April 2026 campaign)

Configuration	Gen Speed	Prompt Speed	Delta vs Baseline	Notes
Dual GPU, all-on-GPU, f16 KV	107 t/s	2,436 t/s	+50% vs single GPU	Final config. TP=2, PCIe x8+x4.
Single GPU, all-on-GPU, f16 KV	71 t/s	—	Pre-dual baseline	After CPU offload fix. Clean VRAM fit.
Full GPU offload (Exp 2 breakthrough)	70 t/s	222 t/s	+118% gen, +200% prompt	vs expert-on-CPU config.
Expert tensors on CPU (MoE routing)	32 t/s	74 t/s	−55%	CPU bottleneck on MoE expert routing.

CPU Inference Baseline — Orchestration Tier (17-day log)

Metric	Value
Sample period	April 2–18, 2026
Model	Gemma 4 26B Q4_K_M
Avg gen throughput	10.49 tok/s
Range	5.02 – 11.72 tok/s
Avg time to first token	616 – 2,224 ms
Avg swap used	4.6 – 7.7 GiB
Speedup vs GPU tier	6.6× (10.5 → 74 tok/s)

Swap usage is the operationally significant CPU-tier finding. In the earlier multi-agent phase, co-locating several always-on processes and local inference on a 32 GB machine consistently drove swap. The architectural fix — separate inference and orchestration nodes — eliminates this entirely. Not a tuning problem; an architectural one.

Autoresearch Results by Category

Optimization delta from experiment baseline (MoE, dual GPU)

full GPU offload (MoE experts)

+118%

dual GPU (TP=2)

+50%

f16 KV cache

+12%

q8_0 KV cache

−18%

expert tensors on CPU

−55%

Agent Infrastructure

Core agent processes run as hardened background services. Auto-restart, post-boot recovery, and local-first inference routing are all part of the runtime design. Slack Socket Mode remains the primary interactive path for the agent stack.

Process	Framework	Interface	Model tier	Status
Agent Harness	Python / custom	Slack Socket Mode	Internal inference gateway (genesis)	live
Mike	RelayV3 (Python)	Discord + Telegram + Slack + IRC	Internal inference gateway	live
Harness Personas	Internal harness roles	Not separately exposed	Internal gateway / task-dependent	internal
Hermes	Python / hermes-gateway.service	Gateway agent — local inference via tower:8010 proxy	Local (genesis INT4 / Nemotron MoE)	live
Chronicle	Python / cron	2am daily cron	Claude API	live

Service Hardening Details

Crash recovery and boot recovery are actively hardened, but the exact unit settings evolve with the live system. The durable design claim is local-first operation: the core agent runtime stays on local infrastructure, while OpenRouter and Claude API remain optional supporting paths.

Raw Data & Artifacts

All benchmark data and logs referenced in the research paper are available publicly.

Artifact	Location	Contents
Live benchmark data	boundarylabs.org/benchmarks.html	Inference metrics, MESA scores, LongMemEval results
Raw experiment logs	inference-research	215+ raw benchmark runs, drivers, and verdict docs per research program
Agent infrastructure	randomchaos7800-hub	Agent harness, benchmark tooling, selected repos
Research paper	papers.html	Full preprint text, tables, references
Weekly build logs	dinovitale.com	Weekly recap posts with build decisions and operational notes