Intelligence Index
- Qwen3.8-27B xhigh34 pts
- GPT-6 Luna High33 pts
- Claude Sonnet 5 xhigh34 pts
APPLIANCE LABS
The right level of intelligence.
At the right size.
The intelligence of GPT-6 Luna and Claude Sonnet 5. On your desk.
Intelligence × size
Independent benchmarkQWEN3.8-27B / EVALUATIONS
Same intelligence class.
Better fit for work.
Qwen3.8-27B is on par with GPT-6 Luna and Claude Sonnet 5 on the Intelligence Index. It stands out especially on tasks that look like a company’s day-to-day work: automating, using tools, working across long contexts and chaining several steps.
AutomationBench
- Qwen3.8-27B xhigh48%
- GPT-6 Luna High48%
- Claude Sonnet 5 xhigh34%
Terminal-Bench 4.0
- Qwen3.8-27B xhigh6%
- GPT-6 Luna High5%
- Claude Sonnet 5 xhigh7%
SciCode
- Qwen3.8-27B xhigh47%
- GPT-6 Luna High50%
- Claude Sonnet 5 xhigh54%
Humanity’s Last Exam
- Qwen3.8-27B xhigh34%
- GPT-6 Luna High33%
- Claude Sonnet 5 xhigh39%
GDP.pdf
- Qwen3.8-27B xhigh17%
- GPT-6 Luna High18%
- Claude Sonnet 5 xhigh13%
CritPt
- Qwen3.8-27B xhigh5%
- GPT-6 Luna High15%
- Claude Sonnet 5 xhigh15%
AA-LCR
- Qwen3.8-27B xhigh82%
- GPT-6 Luna High79%
- Claude Sonnet 5 xhigh77%
Artificial Analysis results, measured independently; these are not Appliance Labs measurements. In bold: best score on each test. Artificial Analysis · accessed October 1, 2026 ↗
THE SYSTEM MAKES THE DIFFERENCE
The model brings the intelligence.
The system turns it into a super intelligence.
Frontier models show how far the state of the art can go. Qwen3.8-27B is the best trade-off we have found to put GPT-6 Luna- and Sonnet 5-class intelligence into a compact appliance.
Around it, we add the context, agents, tools, orchestration and optimizations it needs to actually work for a whole company. And because a task mixes pages to read, small decisions and real reasoning, the Appliance Router hands reading to a Document model and simple decisions to a Decision model: Qwen3.8-27B saves its compute for thinking where it counts.
Hardware + models + harness + orchestration + optimization.
All of it together is what we call Appliance.
The harness helps delegate, keep context, verify and resume long tasks. It does not mechanically raise the model’s Intelligence Index.
HARDWARE
+MODELS
+HARNESS
+ORCHESTRATION
+OPTIMIZATION
↓A SOVEREIGN SUPER INTELLIGENCEFOR THE WHOLE COMPANY
THE LAYERS OF SOVEREIGN SUPER INTELLIGENCE
One Appliance.
A whole system around it.
From hardware to application, twelve layers work together. Explore what turns the model’s intelligence into a sovereign super intelligence for your company.
FROM CHIP TO CONVERSATION
Pick a layer to explore it.
LAYER 01 / 12
Hardware
Power starts in the silicon.
The GB10 pairs a Blackwell GPU with a Grace CPU. Fifth-generation Tensor Cores support FP4, and 128 GB of unified memory holds the model’s weights and its context. Cooling is built into the base.
A hardware foundation designed for local AI compute.
NVIDIA architecture and specs ↗