Evaluate RAM with the workload
Models, context, runtimes and concurrent applications consume memory. T-Chat coordinates available resources; testing on the intended machine remains the way to select a sustainable configuration.
T-Chat for business
Start with the Mac you have.
Grow when you need to.
You do not need the most powerful machine on day one. The right setup depends on the models, the amount of content, the number of simultaneous tasks and the people who will use T-Chat.
Entry16–24 GB
SSD StreamingFrom 64 GB
ScaleUp to 512 GB
T-CHAT 4.6.5
More control in everyday work.
Meetings without loading everything at once
Agents and models do not require a one-to-one relationship. Meeting management accounts for shared models and coordinates heavy local inference.
A dedicated node for the site
For buildings or operational sites, evaluate an Apple Silicon pilot with defined networking, power, access and maintenance. Dedicated enclosures and backup power need partner engineering and validation.
Hardware roadmap
Apple Silicon today.
NVIDIA and AMD in development.
We are developing T-Chat versions for NVIDIA and AMD hardware, extending choice beyond Apple Silicon. They are not available yet: supported operating systems, cards, requirements and release dates will be announced after validation.
The current edition remains designed for Apple Silicon Macs. Memory and model choices should be sized for the actual workload.
Tell us about your hardwareFour levels
Choose around the work you need to do
Unified memory is the most important number for local models. The chip affects speed; memory largely determines which models and context sizes can remain active.
01 · Start
Personal workstation
16–24 GB unified memory
For chat, research, documents, projects and light automation. Smaller local models work well; enabled online models can handle heavier tasks.
- One person or a small pilot team
- Compact local models
- Short or medium contexts
- Online services for workload peaks
Current examples: Mac mini with M6, MacBook Pro with M5, or an existing Apple silicon Mac.
02 · ProfessionalBalanced choice
Daily workstation
32–64 GB unified memory
For heavier daily use, larger documents, more capable local models and more work running at the same time.
- Professionals and small departments
- Medium local models
- Longer contexts
- Smoother automation and AI meetings
Current examples: Mac mini with M5 Pro up to 64 GB; MacBook Pro with M5 Pro.
03 · Advanced
AI department
96–128 GB unified memory
For large models, extended contexts and frequent team use. This tier also gives SSD Streaming more room to work.
- Large local models
- Steady workloads
- More users and processes
- SSD Streaming for supported MoE models
Current examples: Mac Studio with M5 Max up to 128 GB; MacBook Pro with M5 Max up to 128 GB.
04 · Business
Dedicated AI node
256–512 GB unified memory
For very large models, sustained workloads and a central machine dedicated to important business processes.
- Large models fully resident in memory
- More room for context and cache
- More simultaneous activity
- Local, private-network and hybrid use
Current example: Mac Studio with M5 Ultra, configurable with 256 GB or 512 GB.
Grow without disruption
A structure that follows the business
Increase capacity without rebuilding everything. Keep processes, personalities, presets and archives; move only the heavier workloads to a more suitable machine.
One person
T-Chat on the Mac of the first use-case owner. Clear goals, limited data and results that are easy to measure.
One team
Shared configurations, defined roles and smartphone access on the same private network with T-Chat Nearby.
One department
A higher-memory machine for shared local models, documents and automation. Lighter tasks stay on existing workstations.
Several departments
Dedicated nodes for different needs, common rules and online services used only where they create real value.
SSD Streaming
More options, without pretending that storage is RAM
For selected supported MoE models, T-Chat can keep experts on fast SSD storage and load the required ones into memory. This makes some models usable even when they do not fit entirely in RAM, at a lower speed than full in-memory execution.
64 GBpractical baseline for the Flash Q2 profile
100 GB freerecommended minimum on the external drive
Fast SSDa dedicated external NVMe is preferable
SSD Streaming is not universal swap. It works with compatible MoE architectures and still needs memory for the base model, context, cache and active work.
References checked on 29 August 2026: Mac mini, MacBook Pro, Mac Studio and DwarfStar/DS4. Performance varies by model, quantisation, context, cache and simultaneous work. Always validate a configuration against the real use case before purchasing.
Three simple rules
Spend where it changes the work
The most powerful machine is not always the best choice. The balance depends on sensitive data, required speed and the number of people involved.
01
Start with the process
- Which work must stay local?
- How much context is really needed?
- How many people work at once?
- How important is response speed?
02
Use hybrid AI with intent
- Local for data and continuity
- Online for peaks and specialist models
- OpenRouter for a broader model choice
- Clear rules for each workflow
03
Upgrade after the pilot
- Measure speed and quality
- Review memory and cache
- Find the real bottlenecks
- Upgrade only what limits the work
Frequently asked questions
Before choosing a Mac.
Do we need a Mac Studio?
No. An Apple silicon Mac with 16 or 24 GB can be enough for small models, documents and online services. Mac Studio becomes useful as model size, context, sustained load and user count grow.
Does more memory always make T-Chat faster?
Not always. More memory mainly lets you load larger models and keep more context. Speed also depends on the chip, model, quantisation and type of work.
Does SSD Streaming replace RAM?
No. It extends what selected MoE models can do, but it remains slower than unified memory and still requires a substantial amount of RAM.
Can we use T-Chat from a smartphone?
Yes. T-Chat Nearby creates temporary access for devices on the same private network. Companion starts the connection, pairing uses a QR code and access can be revoked.
Can local and online models work together?
Yes. A hybrid setup can keep sensitive workflows local and use selected online services for speed, specialist capabilities or temporary workload peaks.
A configuration built around you
Find the right balance of cost, speed and control.
Start with the computers you already have and the processes that matter. Then build a setup that can grow without waste.