The complete map of open models and hardware.

VRAM (video memory) is the on-card memory that holds a model while it runs. Think of it as a bookshelf: a 30 billion parameter model at Q4 needs about 17-20 GB of shelf space, so a 24 GB card fits it with room to spare, while a 12 GB card does not. Our VRAM fit calculator above computes this exactly for any model you pick.

Quantization shrinks a model's numbers so it takes up less memory, similar to compressing a photo. Q4_K_M, our default recommendation, shrinks a model to roughly a quarter of its full size while keeping about 98% of its quality: the quality drop is not noticeable for drafting, summarising, extraction, Q&A or everyday coding. Try the quantization explainer above to see the size-versus-quality tradeoff yourself.

Most people read at roughly 5 tokens per second, so anything faster feels instant as it streams in. A working example: 40 tokens per second is close to 30 words per second, about one A4 page typed out every 12 to 15 seconds. Use the speed simulator further up this page to watch any model and hardware combination generate at its real estimated speed.

A dense model uses all of its parameters on every response. An MoE (mixture of experts) model, like Kimi K2.6 or Qwen3.6-35B-A3B in the explorer above, has many more total parameters but only activates a small slice for each response, so it needs memory sized to its total parameters but runs at roughly the speed of its smaller active portion. That is why a 1 trillion parameter MoE model can still answer quickly.

Yes, but it needs a multi-GPU server, not a desktop. Per the sizing cheat-sheet above, a 1 trillion-class MoE model needs roughly 550 to 880 GB at Q4, which means an 8x H200 or B200 server (our Rack AI tier). For a single team of 1 to 15 people, a 27B to 70B-class model on a Desk AI or Studio AI build gets you 90% of the practical quality at a fraction of the cost.

Apple Silicon's unified memory lets one pool of RAM serve as VRAM, so a 128 GB or 256 GB Mac Studio can load huge models a same-priced GPU cannot fit. NVIDIA GPUs are faster per GPU on raw throughput and have the widest software support (CUDA, vLLM). Rule of thumb: choose Apple Silicon to fit the biggest model on one quiet box; choose NVIDIA for the fastest single-user response or when you need vLLM-grade multi-user serving.

It depends on the license, not the fact that it is "open." Apache-2.0 and MIT models (Qwen3, DeepSeek, GPT-OSS, GLM) are free for commercial use with no per-token fee ever. A few models, like Mistral Large 2 or Codestral, carry research-only or non-production licenses that need a separate commercial agreement. See the licensing table above; VYROX defaults every build to permissively-licensed models so you own the deployment outright.

A one-GPU Desk AI build draws about the same power as a gaming PC and sips a few ringgit of electricity a day on a standard wall socket. An 8-GPU Rack AI server draws roughly 6 to 7 kW, needs 3-phase power and real room cooling, and is loud enough that it belongs in a server room, not an open office. The power and operations table above breaks down every tier in between.

Match headcount first, model size second: 1 to 3 people generally fits Desk AI, 5 to 15 fits Studio AI, 15 to 40 with background agents fits Engine AI, and a whole organisation fits Rack AI. See the "Which tier fits your team size" decision helper above for the exact hardware and price behind each, or use the "what can I run on my machine" tool if you already own hardware.

No. Because you own the hardware and the deployment stack, swapping in a newer open model, such as whatever supersedes Qwen3.6 or DeepSeek V4, is normally a download and a config change rather than a new purchase. The GPU or Mac you buy today keeps running whichever model is best for years; only the very largest frontier-scale models eventually ask for more memory than a given card holds.

Work it out from the peak draw of your tier and your own tariff. As an illustration only, at RM 0.55 per kWh, a Studio AI build held at its 0.9 kW peak for 10 hours a day across 22 working days comes to 198 kWh, roughly RM 109 a month. That is a deliberate worst case, because a real machine idles between prompts rather than sitting at peak. The power and cooling section above works the same calculation for every tier so you can substitute the rate on your own bill.

Not for the smaller tiers. Desk AI sits on a normal desk on a standard 13A socket, and Studio AI wants a dedicated 13-15A circuit and a ventilated corner rather than a sealed cupboard. Engine AI is where it becomes a facilities question: a 20A circuit, a UPS and a cupboard with a door, because the fans are server-grade. Rack AI needs a genuine server room with 3-phase power and room cooling. We survey the site and confirm all of this before any hardware is ordered.

Answer four questions in order: what the work mostly is (general office work, coding, or reading scanned documents), which languages it must handle, how long the longest document you feed it is, and how much memory you have or intend to buy. The first two decide the model family, the last two decide the size. The "Choosing a model in four questions" section above walks through each one, and we test the shortlist on a sample of your own real documents during the audit rather than trusting a leaderboard.

Electricity every year at your own tariff, the optional managed-service tier if you want monitoring and upgrades handled for you, and a service and parts allowance from year three when manufacturer warranty on workstation and datacentre parts typically ends. What is not in the budget is just as important: no per-user licence fees, no per-token charges and no subscription renewals, because Apache-2.0 and MIT open-weight models carry none of those. The three-year section above breaks this down line by line.

Sizing for the model weights and forgetting the context memory. A 32B model at Q4 looks like it fits a 24 GB card until the first long contract is loaded and the KV-cache pushes it over. The other frequent errors are buying an edge accelerator that cannot run a language model at all, defaulting to multi-GPU when one larger-memory card would be simpler, checking the model licence only after deployment, and specifying the machine without specifying the power and ventilation around it.