Holon Research, Inc

Infrastructure for private intelligence.

Reserve beta unit

From $20,000 · 8 units left

Holon inference computer

Plug and Play inference computer

Custom engine

Up to 3x the performance of standard engines.

Air-gapped inference

No cloud, no telemetry.

8-16 users

Serve your whole team on your LAN.

Uncensored & abliterated models

Your weights, your rules.

Multi-user stack included

Accounts, queues, serving on day one.

Fine-tune from usage

Traces become training data.

< 5 minute setup

Outlet to first token.

Custom engine

Up to 3x the performance of standard engines.

Air-gapped inference

No cloud, no telemetry.

8-16 users

Serve your whole team on your LAN.

Uncensored & abliterated models

Your weights, your rules.

Multi-user stack included

Accounts, queues, serving on day one.

Fine-tune from usage

Traces become training data.

< 5 minute setup

Outlet to first token.

Holon private inference computer

Custom C++/HIP engine designed for 4×R9700. Squeezes every drop of compute.

Up to 16 Concurrent users
Up to 20 Billion Tokens per month
Max 1,400W Power draw
Inside the Holon inference computer

Based on Qwen3.8-27B MXFP4, 90/10 input/output split, 99% cache hit rate, and 60% utilization over 30 days.

Model profile Quant KV type Users Prefill Peak aggregate decode at 4K Context window
Qwen3.8-27B MXFP4 MXFP4 BF16 16 9,830.96 tok/s 1,426.02 tok/s 131,072 tokens
Qwen3.8 Next Flash Int4/FP8 Int4/FP8 BF16 16 7,243 tok/s 849 tok/s 131,072 tokens
Qwen3.8-27B FP8 FP8 BF16 8 4,453 tok/s 322.26 tok/s 139,264 tokens
DeepSeek V4.0 Flash 0731 native FP8 4 2,268 tok/s 105.99 tok/s 262,144 tokens

Supported models

  • Qwen3.8-27B
  • Qwen3.8 Next Flash
  • DeepSeek V4.0 Flash 0731

Uncensored / Abliterated

  • Qwen3.8-27B Heretic MTP FP8
  • Qwen3.8-Flash-Next OrcaUncensored IQ2-Light
  • DeepSeek V4 Flash 0731 Abliterated
  • GLM 5.3 Flash Uncensored/Abliterated

Need to deploy a custom model instead?

Reach out

< 5 minute setup

Plug into any standard 15-amp outlet, assign IP, serve inference

Embed nearly 250,000 tok/s for RAG

Supports: pplx-embed-context (BF16), Nemotron-3.5-Embed-1B-BF16

Unlimited web search,
unlimited deep research

Use self-hosted tools, or BYO

Train your specialized model

Holon inference computer front face
Inside the Holon inference computer

Get Holon

  • GPUs 4 × AMD Radeon AI PRO R9700 32GB GDDR6 128GB VRAM
  • CPU AMD EPYC 5th Gen 9005 Series Twenty-four core · Model 9255
  • Memory 128GB DDR5 ECC 5600 MT/s
  • Power 1600W Titanium PSU 80 PLUS Titanium
  • Storage 4TB M.2 SSD NVMe
  • Stack Holon's engine, runtime environment, training stack
Save $5,000 during beta US shipping only
$25,000 $20,000

Beta price

8 units left in the beta batch · 2 of 10 booked

Reserve beta unit ($1,000)

Ships week of Dec 1st ·

Request remote access

Try a unit before you buy