PRICING
CurXor is $3,999 once for the appliance. Economy, Balanced, and Performance are not SaaS plans; they are recommended UMA allocations for local models on the same 64 GB system.
CURXOR STANDARD 64
$3,999once
One-time hardware · no CurXor subscription · $0/mo CurXor API on the operate plane
Capital, Creator and Outreach are demo-ready desks; eight further crewmates ship as labelled previews. Outbound stays simulated until you wire your own bridges on eno2.
Pre-order reserves your place in the queue. Ship dates are announced by email as production batches are scheduled — no delivery date is promised at checkout. Pre-order terms
OPEN-WEIGHT COMPUTE LADDER
CurXor OS — mint autonomous employees on bare metal. Standard 64 GB today. Pro and CurXor Studio tomorrow. Same sovereignty. Stronger models as memory allows. Press kit.
| TIER | Standard$3,999 · 64 GB UMA · pre-order | Pro 128$4,999 indicative · 128 GB UMA · roadmap · no checkout | CurXor StudioTBD · 256 GB+ UMA · horizon · no checkout |
|---|---|---|---|
| PLATFORM | MS-S1 Max · Ryzen AI Max+ 395 | Same platform · 128 GB SKU | CurXor-branded custom box |
| DEFAULT LOCAL STACK | moondream:1.8b + qwen3.5:9b | Above + qwen3-vl:8b · qwen3:14b · qwen3.6-35b:q4 | Largest open-weight models at usable quant |
| BEST FOR | Flagship desks — 3 demo-ready · daily operate | Power operators — heavier local models · dual hot-load | Horizon — largest open weights when memory demands it (no pre-order) |
| OPERATE INFERENCE | $0/mo on appliance | $0/mo on appliance | $0/mo on appliance |
Honest footnote: $0/mo operate inference = local LLM on the appliance for the operate plane (chat, planning, Forge assist). Does not include optional third-party BYOK, trading/social bridges, or your power bill.
Directional evidence (same silicon class as MS-S1): pi-local-coding-bench.dev — Strix Halo / same silicon class evidence — directional, not a guarantee on your SKU until validated.
Standard 64 is the only tier you can buy today. Pro 128 pricing is indicative and CurXor Studio is post-traction horizon — neither has a checkout. Press ladder copy
COMPUTE LADDER
UMA tiers below are model budgets on CurXor hardware — not separate products.
CLAWBOX / JETSON CLASS
8 GB
~67 TOPS NPU class
Light always-on assistants
MAC MINI / DIY OPENCLAW
16–32 GB
Varies NPU class
You assemble the stack
CURXOR STANDARD 64
64 GB UMA
126 TOPS NPU class
Turnkey wealth · work · life crewmates
EVOLVING MODEL
Hardware purchase is the anchor — $3,999 once, $0/mo operate-plane API on the appliance. CurXor does not bill per token for local inference or bundled crewmates.
CurXor OS evolves on metal you already own via OTA: compute tiers from Standard to Pro 128 to CurXor Studio as open weights advance, crewmate depth unlocks as flagship desks mature, and The Forge lets you mint custom operators without a new subscription per workflow.
Frontier models and bridge connectors stay optional BYOK — additive, never required to run the desk. We are not building another API rent trap.
MODEL BUDGETS BY UMA TIER
ECONOMY
Minimal UMA footprint · moondream + qwen3.5:9b · single-crewmate kiosks
vision · 4 GB UMA · Tiny vision-language for kiosk cameras and lane vision.
reasoning · 7 GB UMA · Default agent backbone — tool-calling + thinking on gfx1151 UMA.
reasoning · 4 GB UMA · Economy/fast reasoning — smaller UMA footprint when 9B is overkill.
BALANCED
Recommended for most crewmates · day-one moondream + qwen3.5:9b (VL/14b only if pulled)
vision · 4 GB UMA · Tiny vision-language for kiosk cameras and lane vision.
vision · 8 GB UMA · Optional stronger vision (~6 GB) — 256K context; keep moondream as day-one default (swap cost).
reasoning · 7 GB UMA · Default agent backbone — tool-calling + thinking on gfx1151 UMA.
reasoning · 4 GB UMA · Economy/fast reasoning — smaller UMA footprint when 9B is overkill.
reasoning · 11 GB UMA · Stronger multi-step planning for content and capital desks.
vision + action · 18 GB UMA · Robotics VLA — vision + action head for manipulation.
PERFORMANCE
Max quality on 64 GB UMA · qwen3.6 MoE for coding · VLA when manipulation intent
vision · 4 GB UMA · Tiny vision-language for kiosk cameras and lane vision.
vision · 8 GB UMA · Optional stronger vision (~6 GB) — 256K context; keep moondream as day-one default (swap cost).
reasoning · 7 GB UMA · Default agent backbone — tool-calling + thinking on gfx1151 UMA.
reasoning · 11 GB UMA · Stronger multi-step planning for content and capital desks.
reasoning · 22 GB UMA · Official Ollama MoE — strong general agent backbone on 64 GB UMA.
reasoning · 20 GB UMA · Strix Halo coding leader — ~67% SWE-mini on pi-bench, not yet measured on MS-S1; agentic tool-calling.
reasoning · 28 GB UMA · Higher-quality Qwen3.6 quant for Pro 128 GB coding / heavy agent work.
reasoning · 18 GB UMA · Dense coding model — pi-bench sweet spot when 128 GB allows hot-load headroom.
reasoning · 24 GB UMA · Dense flagship reasoning for Pro 128 GB hot-loaded stacks.
vision + action · 18 GB UMA · Robotics VLA — vision + action head for manipulation.
Everything without a badge runs on the Standard 64 appliance you can buy today. Entries marked NEEDS PRO 128 are catalog headroom for the 128 GB SKU — listed so you can see where the stack goes, not included at $3,999.
BEYOND HARDWARE
CurXor is a capital expense once — compare cloud rent, assistant subscriptions, and other paths with visible assumptions on the illustrative spend scenarios table.
WHAT CHANGES
Model quality, vision depth, and whether you budget enough UMA for VLA workloads.
WHAT DOESN'T
The hardware price, Home/Settings/Forge UI, and the option to stay fully local by default.
OPTIONAL CLOUD
Frontier models remain bring-your-own-key. They are optional and not part of CurXor pricing.
ONE PRICE, ONE BOX
$3,999 once for the appliance. Everything above is a memory setting, not another invoice.
