FAST-CHAT · AGENT
Rigel
The fast-chat tier and the always-on agent’s brain. Dual RTX 3090, one of them on a custom water loop.
48GBWin 11
Robinson Labs / Overview
An independent AI lab that runs its own models.
Seventeen GPU nodes between a garage and an office, two NAS boxes, a custom water loop, and a 40-gigabit RDMA llama.cpp cluster. Roughly four million tokens of live context across the serving fleet. Local inference, persistent memory, no rented GPUs. This is the build log.
17 GPU NODES · 40G RDMA FABRIC
NVIDIA (Orion)AMD (Vela)StorageSupport · no GPULines = RPC / topologyBrightness = health · as of 27 Oct 26
SCROLL
Outcomes
DELIVERABLES · AS OF 26 Oct 26
347
Model downloads, last 30 days
2
Open models on Hugging Face
93%
Merged PRs agent-authored
3997 of 4291 merges across 34 own-org repos. Every merge human-reviewed.
How these are counted
Downloads are Hugging Face rolling 30-day totals across published Robinson Labs models, a public counter I do not control. A merge is agent-authored when the PR body carries the Generated-with-Claude-Code trailer left by the pr-architect flow. Counts cover Robinson Labs own-org repositories only, and every merge is human-reviewed before it lands, so an agent opens the PR but a person makes the call. Numbers refresh at deploy time.
Cluster
Two constellations, one sky.
17 GPU NODES · 40G RDMA FABRIC
ORION
orion cluster
The NVIDIA RPC constellation: six nodes and 16 GPUs on 40G RDMA, pooling about 150 GB of VRAM behind one coordinator.
150GBUbuntu
FLEET
Pleiades
10× Pascal cards for the embed / rerank fleet, plus the Whisper STT lane.
~86GBUbuntu
FAST-CHAT · POOL
Alsephina
Active member of the fast-chat pool, serving the custom 35B behind the load balancer. 3× RX 5700 XT on Vulkan. 128K × 3 context.
24GBUbuntu
VISION
Suhail
4× RX 5700 XT on Vulkan. Vision summarizer, Qwen3.6-35B-A3B.
32GBUbuntu
IMAGE-GEN
Regor
SDXL + Lightning on sd.cpp/Vulkan, four cards driven from Rigel. 4 images per pass, ~20s a pass.
4×8GBUbuntu
INSTRUCT
Markeb
Structured-JSON lane: 2× Qwythos-9B pool, one per card. Clean JSON, clean tool calls.
16GBUbuntu
AGENT-BRAIN
Avior
3× RX 5700 XT on Vulkan. Agent-brain lane serving the AgentWorld-35B quant that won the tau-bench eval. Brightest star of the False Cross.
24GBUbuntu
AGENT-BRAIN
Aspidiske
3× RX 5700 XT on Vulkan. Second AgentWorld-35B agent-brain lane. With Avior, Alsephina and Markeb it completes the False Cross.
24GBUbuntu
AGENT-BRAIN
Acrux
3× RX 5700 XT on Vulkan. Third AgentWorld-35B agent-brain lane. Brightest star of the real Southern Cross, the one the False Cross gets mistaken for.
24GBUbuntu
INSTRUCT
Naos
Second Qwythos-9B instruct pool and failover twin for Markeb, so either box can drop without taking the structured-JSON lane down.
16GBUbuntu
STORAGE
Altair · Deneb
Two-NAS topology. Media archive, model library, and the dual-Xeon build host.
2× NASunRAID
Orion star assignments are illustrative; firm names land at each rebuild. Avior and Aspidiske are firm: with Alsephina and Markeb they complete the False Cross. Acrux sits just outside it, the brightest star of the real Southern Cross that the False Cross gets mistaken for. Naos is in Puppis, another piece of the same ship. Constellation names are locked.
Models in production
Four brains, one fleet.
~4M TOKENS LIVE CONTEXT · SERVED LOCALLY
Rigel · fast-chat / agent
Fast-chat slot
quant picked per residency
48GBVRAM · DUAL 3090
Orion cluster · deep
Cluster-served model
quant picked per residency
150GBVRAM · 16 GPUS
Alsephina · fast-chat pool
Qwen3.6-35B-A3B
custom domain quant
384K128K × 3 CTX
Avior · Aspidiske · Acrux · agent
Qwen-AgentWorld-35B-A3B
MTP graft · ~75% draft acceptance
512K256K × 2 CTX · PER RIG
Markeb · Naos · instruct
Qwythos-9B × 4
~190TOK/S POOLED
Suhail · vision
Qwen3.6-35B-A3B
35BMOE VISION
Regor · image-gen
SDXL + Lightning
~20sPER PASS · 4 IMG
Voice · STT + TTS
Whisper large-v3-turbo · Kokoro-82M · Piper
3VOICE LANES
Pleiades · retrieval
Qwen3-Embedding · Qwen3-Reranker
4EMBED + RERANK
Published to Hugging Face
A snapshot of the repos the model pipeline has shipped, with the provenance chain that built each one. Generated from the pipeline catalog, as of 27 Oct 26.
Qwen3.5-122B-A10B-REAP-30-abliterated-GGUF
Abliterated from Qwen3.5-122B-A10B-REAP-30, then converted, then quantized.
Qwen3.5-122B-A10B-REAP-30 (base) → abliterated → BF16 → Q8_0
Q8_0 · 93 GBQ6_K · 72 GBQ5_K_M · 62 GBQ5_K_S · 60 GBQ4_K_M · 53 GBQ4_K_S · 50 GBIQ4_XS · 47 GBQ3_K_M · 42 GBIQ3_M · 39 GBQ3_K_S · 38 GBIQ3_XS · 36 GBIQ2_M · 29 GBIQ2_XS · 26 GB
06 Sep 26
Qwen3.5-122B-A10B-abliterated-GGUF
Abliterated from Qwen3.5-122B-A10B, then converted, then quantized.
Qwen3.5-122B-A10B (base) → abliterated → BF16 → Q8_0
Q8_0 · 133 GBQ6_K · 102 GBQ5_K_M · 89 GBQ5_K_S · 86 GBQ4_K_M · 76 GBQ4_K_S · 72 GBIQ4_XS · 68 GBQ3_K_M · 61 GBIQ3_M · 56 GBQ3_K_S · 55 GBIQ3_XS · 52 GBIQ2_M · 42 GBIQ2_XS · 38 GB
06 Sep 26
Qwen3.6-35B-A3B-abliterated-GGUF
Abliterated from Qwen3.6-35B-A3B (bf16, MTP base), then converted, then quantized.
Qwen3.6-35B-A3B (bf16, MTP base) (base) → abliterated → BF16 → Q8_0
Q8_0 · 38 GBQ6_K · 29 GBQ5_K_M · 25 GBQ5_K_S · 25 GBQ4_K_M · 22 GBQ4_K_S · 21 GBIQ4_XS · 19 GBQ3_K_M · 17 GBIQ3_M · 16 GBQ3_K_S · 16 GBIQ3_XS · 15 GBIQ2_M · 12 GBIQ2_XS · 11 GB
03 Sep 26
Qwen3.8-27B-abliterated-GGUF
Abliterated from Qwen3.8-27B (bf16 base), then converted, then quantized.
Qwen3.8-27B (bf16 base) (base) → abliterated → converted → Q8_0
Q8_0 · 29 GBQ6_K · 22 GBQ5_K_M · 20 GBQ4_K_M · 17 GBIQ4_XS · 15 GBQ3_K_M · 14 GBIQ3_XS · 12 GBIQ2_M · 10 GB
19 Aug 26
Qwen3.5-REAP-262B-A17B-abliterated-GGUF
Abliterated from OpenMOSE/Qwen3.5-REAP-262B-A17B (bf16 safetensors, community base), then quantized.
OpenMOSE/Qwen3.5-REAP-262B-A17B (bf16 safetensors, community base) (base) → abliterated → Q5_K_M
Q5_K_M · 186 GBQ4_K_M · 159 GBQ4_K_S · 149 GBIQ4_XS · 140 GBQ3_K_M · 125 GBIQ3_M · 115 GBIQ3_XS · 107 GBIQ2_M · 86 GBIQ2_XS · 77 GBIQ2_XXS · 69 GB
18 Jul 26
About
A one-person lab with a real cluster.
INDEPENDENT · SELF-FUNDED · NC
Robinson Labs is the homelab of Stephen Robinson, a ServiceNow architect in Charlotte, NC who rents frontier intelligence and owns the rest. By day I design enterprise platforms; by night I run a local-first AI operating system on hardware I own.
The bet is simple: own your models, your memory, and your tools. Local inference, persistent memory across sessions, agentic tooling that does real work, no rented GPUs, no API key between a request and the models I serve. The interesting problems live in the wiring, not the billing.
The cluster started as a pile of PCs under my desk; today the rack lives in the garage on a lit 40-gigabit RDMA fabric. This is where I write down what I learn building it: the dead ends, the BIOS resets, the “nothing changed but everything broke” nights.
LocationCharlotte, NC
GPU nodes17
RPC cluster VRAM~150 GB · 16 GPUs
Fleet VRAM400 GB+
Fabric40 GbE RDMA (RoCE)
Live context~4M tokens
StorageTwo-NAS · unRAID
Stackllama.cpp · CUDA · Vulkan
RuleLean. No bloat.
Stack
The work on top of the metal.
NAMES ABSTRACTED · PRE-RELEASE
W/01
Distributed AI platform
Agents, tools, chat: the spine everything else plugs into.
W/02
TCG collection assistant
Singles, sealed + graded subsystem, open-a-box as an atomic transaction.
W/03
Crafting assistant
AI-assisted CRUD for a real working maker. Daily-driven.
W/04
Persistent memory + kanban
The lab's long-term memory: decisions, facts, work items across every project.
W/05
Model-manufacturing pipeline
Abliterate, prune, distill, quantize, with a provenance DB behind it.
W/06
PC-build configurator
Live GPU scraping + a multi-pass enrichment pipeline.
W/07
Agentic social platform
Early build. Agent-driven content workflow.
Early
W/08
ServiceNow MCP + agentic SDK
A large MCP tool surface plus an SDK-native agentic layer. Personal capability-build.
Parked
Writing
What we ship, we write down.
ENGINEERING JOURNAL · WAR STORIES
25 Oct 26
War story · Ops
The freeze was a snapshot every five minutes
rigel hung for up to a minute whenever agents were busy, and the game on it kept crashing. We priced a read cache before we found the poller that was snapshotting the whole C: drive about 288 times a day.
5 min
18 Oct 26
War story · Lab
The BF16 file that was not BF16
We asked the converter for bf16 and named the file after our request. What came out was 96 percent MXFP4 by parameter count, and it cost us a failed quantize before anyone looked inside.
5 min
04 Oct 26
War story · Cluster
One bad prefill, twelve failures
A replay said our newly promoted serving line collapsed on long prompts in 11 of 12 runs. A rollback was queued. The twelve samples were really two.
5 min
27 Sep 26
War story · Cluster
The probe that cost 217 watts
Every GPU in the cluster woke up once a minute with nobody using it. The culprit was our own health check: a one-token request that took 0.65 seconds of compute and held sixteen cards at raised power for twenty.
6 min
20 Sep 26
War story · Ops
The cable was fine
A new server's 40-gigabit link fell back to 10 after every boot. We blamed the cable, then autonegotiation, then firmware. The answer was a BIOS setting, and the test that found it was free and took two seconds.
6 min