Adaptive memory management
Resource management considers machine memory and workload. Sizing remains specific to the model, context and tasks; it does not require one fixed hardware configuration.
AI infrastructure
Choose the model. Understand the machine.
Choose local or online models and see the real state of the Mac. T-Chat shows what is ready, what uses resources and which limits matter.
RuntimesMultiple
StatusReal
SSDStreaming
T-CHAT 4.6.5
More control in everyday work.
Parameters and readable responses
Supported model parameters can be adjusted through the available workflow. Fixes in the 4.6 series also improve character handling in compatible runtime responses.
Visible Hugging Face downloads
Companion shows status, transferred amounts and percentages when the total is known. Error details help distinguish download issues from runtime prerequisites.
A local control layer
Multiple engines, one experience
Ollama, Hugging Face models, ComfyUI and specialised runtimes appear in the interface when they are genuinely ready.
Models
Open weights without lock-in
- Ollama models available in chat
- Hugging Face repositories running through MLX or llama.cpp
- Qwen and Gemma parameters adjustable during chat
- Context and tokens measured against the real limit
Companion
The machine remains understandable
- Status for Docker, database, Redis, Ollama and ComfyUI
- Start, stop and restart services
- Model observatory using real memory and request data
- T-Chat Nearby with QR pairing and revocable private-network access
Expansion
Beyond available memory
- External SSD configuration for DwarfStar
- SSD Streaming for experts in supported MoE models
- Local engines for images and video
- Clear separation between installation and active model
Open optimisation
T-Chat is built for open-weight models
The experience is optimised for the Qwen family and Google Gemma without locking the platform to one provider.
Live parameters
Temperature, top-p, top-k, min-p, seed and context can be adjusted for models that expose them.
Real observability
Charts and status come from observed processes, memory and requests rather than decorative animation.
External models and TimesFM 3
Ready local models are registered through Companion. TimesFM 3 is activated separately as a forecasting tool for personalities, chats, meetings, missions and Automata.
SSD Streaming
Supported MoE models keep experts on fast storage and load the required ones into memory. It is slower than RAM and is not universal swap.
Build your AI perimeter
Use the right models without losing infrastructure visibility.
Let's assess the machine, runtimes and requirements for a sustainable local configuration.