GPU
NVIDIA RTX 5090 · 32 GB
Every token is computed here, on a single consumer graphics card.
Private AI · one machine
Private
intelligence.
A 27-billion-parameter open model, running on one machine I own. Chat, agent and long-context research — invitation only.
- GPU
- RTX 5090
- Model
- Qwen3.8 · 27B
- Decode · bench
- 110 tok/s
01 · The machine
One card.
No cloud model.
Model
Qwen3.8 27B · NVFP4
A 27-billion-parameter open model, quantised to 4-bit so it fits one card with room for long context.
Speed
Speculative decoding
A small draft model guesses several tokens ahead; the large model checks them in one step.
02 · Three modes
Three ways to think.
Chat
Thinks, then answers. Fast back-and-forth, with its reasoning there when you want to see it.
Agent
Uses tools. Searches the web, opens pages and works through multi-step tasks on its own.
Deep
Long-context research. Reads and extracts from many sources, then writes only what it observed.
03 · Private by design
The model runs
on my hardware.
- No third-party model API. The model itself runs on one machine I own.
- Encrypted in transit. Every request travels over TLS to a private gateway, where your chat history is kept.
- Never used for training. Your chats are not sent to any model company.
- Only search queries leave. When Agent or Deep searches the web, the query goes to a search service — never the conversation.
04 · Measured
Numbers, not adjectives.
- decode speed
- 110tok/s
- warm first token
- 65ms
- context window
- 114K tokens
- graphics memory
- 32GB
Benchmarked on 22 Aug 2026 · speculative-decoding profile · 1K-token generation, single user.
05 · Access
Invitation only.
MohebiAI is a private workspace for a small circle. There's no public sign-up.
Sign in