Source: https://tensorquay.com/ LOCAL AI WORKSTATIONS / SINGAPORE # Local AI. Built for your work. RTX 5090 AI workstations, local models and agents. Configured together for individuals and teams in Singapore. YOUR TASK “Summarise a report.” Illustrative workflow Tools and agent setup agreed for your project. [Explore the AI Station](https://tensorquay.com/#platforms) TQ–32 · currently not available S$15,615 PRODUCTS / AI STATIONS ## Your AI Station. Ready to work. Local models and an agent, preinstalled. Configured for your workflows. 01 / INDIVIDUALS & SMALL TEAMS Not available ### TQ–32 Local intelligence. Ready to work. Qwen 27B and an AI agent, installed and configured for your work. PRICE / SGD S$15,615 Image: NVIDIA GeForce RTX 5090 Founders Edition GPU; official component photography, not the complete workstation NVIDIA GeForce RTX 5090 NVIDIA component image THE COMPLETE SETUP #### More than the machine. Your task Local AI Your result - Qwen 27B installed - Agent configured for agreed tasks - Documentation & guided handover [Explore AI deployment](https://tensorquay.com/local-ai-deployment.html) GPU RTX 5090 32 GB VRAM · FE or board-partner card MEMORY 64 GB DDR5 RAM STORAGE 2 TB NVMe SSD LOCAL AI Qwen 27B Model + agent setup Workstation. AI setup. Guided handover. [Full specifications](https://tensorquay.com/hardware.html#tq32) Image: Illustration of one agent server routing work to several agents ### TQ–A128 Available to order · delivered within 1 week Agent server for teams · 16 cores / 32 threads · 128 GB RAM · no GPU needed TensorQuay Agent set up for your team: user accounts and roles, a sandbox per agent task and a model router to local or cloud models. Model-agnostic. 3-year parts warranty, labour included · carry-in service S$8,490 · assembly, burn-in, agent setup and guided handover included · [Full specifications](https://tensorquay.com/hardware.html#tqa128) Image: NVIDIA RTX PRO Blackwell professional GPU; official component image, final system subject to confirmation ### TQ–192 Proposed · by enquiry For larger local models and team workloads, with 192 GB of GPU memory. Proposed; available by enquiry. Proposed 2 × RTX PRO 6000 Max-Q · 192 GB GPU memory From S$69,900 (Hardware + Ubuntu) · S$74,900 with AI setup · ordered by quotation SEE IT AT WORK RECORDED TESTS / 02 SEP 2026 ## One GPU. See it work. Qwen 27B. One RTX 5090. Real agent sessions, running side by side. 8, 16 & 30 SESSIONS · ORIGINAL-SPEED FOOTAGE SCREEN RECORDING 8 SESSIONS 60 SECONDS · NO AUDIO [Open the 8-session test video](https://tensorquay.com/assets/benchmarks/qwen27b-5090-8-sessions.mp4) . CHOOSE A RECORDING ### Eight tasks. One shared model. Agents work on separate coding tasks: reading files, making edits and running checks. Each terminal is a separate session. [View full-size](https://tensorquay.com/assets/benchmarks/qwen27b-5090-8-sessions.mp4) Internal tests on a rented RTX 5090 cloud system. These clips show the agent in use; performance varies with the task and conversation length. [Read the test results & limits](https://tensorquay.com/rtx-5090-local-ai-benchmark.html) Explore the measurements & test conditions Recorded agent work and short-request throughput are different tests. The figures below come from scripted requests, not from timing these videos. Model Qwen3.8-27B · NVFP4 GPU 1 × RTX 5090 · 32 GB VRAM Runtime vLLM 0.28 · FP8 KV cache ### What the recordings show Separate agents sharing one model server while editing and testing small programs. The visible interface is our engineering test harness. The 16-session run completed 16 tasks in the test log, with a median of 7.5 minutes. At 30 sessions, 27 tasks completed, two failed and one was still running when observation ended. The median for completed tasks rose to 20.7 minutes. Thirty sessions exceeded practical capacity for this workload. A session can be waiting or calling a tool; it does not mean all sessions are generating tokens at the same instant. Clips are continuous, silent excerpts at original speed: 8 sessions, 00:15–01:15; 16 sessions, 00:45–01:45; 30 sessions, 15:00–16:00 within the source recordings. No tasks are removed from the selected frames. SHORT-REQUEST THROUGHPUT RTX 5090 / 32 GB 445 TOK/S APPROX. COMBINED OUTPUT 57 TOK/S APPROX. PER SESSION, MEASURED 8 sessions: recorded short-request baseline. Short scripted requests, up to 512 output tokens. Per-session speed is measured over each request’s own start-to-finish time; it is not the combined output divided by the number of sessions. The 32-session measurement is separate from the 30-session recording. These tests do not establish a supported customer count or guarantee complete workstation performance. LOCAL VIDEO GENERATION / RTX 5090 ## From a prompt. To a scene. A hawker story. An English presenter. Mandarin dialogue. Explore video and sound generated together with FastH3 V2 on one RTX 5090. AI-GENERATED · MINIMAX H3 FASTH3 V2 / 8 STEPS Samples are available in permitted regions. Open a video below, or enable JavaScript to use the player. [Open the hawker story video](https://api.tensorquay.com/api/video-generation/fasth3-5090-hawker-story.mp4) . Selected generated outputs, at original speed with their generated audio. These are samples to inspect, not recordings of the generation process. OUR TEST / 18 SEPTEMBER 2026 ### A little Singapore story. A hawker hall, a cook at the wok, then a shared meal. Three generated shots with audio in one 15.1-second clip. Clip length 15.1 seconds Generation time 219.4 seconds Generated at 1344 × 768 Web playback 1344 × 768 Watch the scene changes and the stall signs. The atmosphere comes through; the lettering is unreliable. [Open video](https://api.tensorquay.com/api/video-generation/fasth3-5090-hawker-story.mp4) to generate a 5.2-second clip 56.6 seconds to generate a 10.1-second clip 126.5 seconds to generate a 15.1-second clip 218.2 seconds Warm medians across four prompts at 1344 × 768, including audio and file saving. FastH3 V2 INT8, 8 steps, sparse attention; one RTX 5090 cloud GPU with 32 GB VRAM and 125 GB host RAM. These tests do not qualify the complete TQ–32 or concurrent Qwen and video workloads. [Results, samples & limits](https://tensorquay.com/rtx-5090-video-generation.html) MiniMax H3 samples are unavailable in the US, EU, UK and South Korea, or where the viewing region cannot be established. [Licence terms](https://tensorquay.com/licences/minimax-h3.txt) · [IP Geolocation by DB-IP](https://db-ip.com/) . YOUR WORKSTATION. OUR SUPPORT. AFTER-SALES / WARRANTY & SUPPORT ## A little care. A long way. One team for hardware diagnosis, warranty claims and your local AI setup. Hardware warranty, parts & labour 3 years Planned burn-in testing 72 hours Optional software maintenance S$299 / year [Explore support](https://tensorquay.com/solutions.html#support) Coverage & service details Every TQ–32 has a 3-year parts and labour warranty on the workstation hardware. TensorQuay handles warranty claims, so you contact one company. Service is pickup and return in the first year, then carry-in service. The warranty does not cover accidental damage, misuse, liquid damage or unauthorised modifications. Planned burn-in checks cover the GPU, memory, storage and AI setup under sustained load before delivery. Software maintenance is optional and not required to use your workstation; without it the machine keeps working exactly as delivered. It includes model and runtime updates, driver and security patches, and email support. Remote sessions require your permission and are time-limited; there is no standing access. The maintenance scope, response times and number of remote sessions per year are being confirmed; discuss these with us before choosing a support plan. On-site hardware repair is not offered. Repair time depends on parts availability. Ask us for the full warranty terms before ordering. Ongoing AI changes and software support will have a separately defined scope. CONFIGURED BEFORE DELIVERY ## Your work. Our starting point. Models installed. Workflows configured. A guided handover to get you started. - 01 Agree the scope - 02 Configure & test - 03 Hand over to you [Explore deployment](https://tensorquay.com/local-ai-deployment.html) Remote deployment · On-site in Singapore [Need a place for your computer? Explore Singapore hosting](https://tensorquay.com/solutions.html#hosting) What’s included Choose the tasks, documents and integrations, and agree what a successful setup looks like. We install the model and runtime, configure your agent and check the agreed tasks alongside the hardware. You receive your configured system, documentation and a guided walkthrough of your workflows. Project scope and access arrangements are agreed first. On-site deployment is available in Singapore only. [Support & model adaptation](https://tensorquay.com/solutions.html) MODELS RUNTIME AGENTS BEFORE YOU START ## A few good questions. The details behind the solutions. Can I order today, and how long does it take? The TQ–32 is currently not available to order. Contact us and we will let you know when it is. What is the TQ–A128, and how fast is delivery? The TQ–A128 is our agent server for teams: 16 cores / 32 threads, 128 GB RAM, no GPU needed. We set up TensorQuay Agent, our open, on-premise agent platform (pre-release), for your team. It is available to order and delivered within 1 week , within Singapore. Included and set up for you: user accounts and roles; a sealed sandbox per agent task; a model router to local or cloud models (any OpenAI-compatible endpoint); secrets kept outside the sandboxes; a network allow-list; shared workspaces; tool connectors (MCP); and a shared channel where agents and people coordinate. Coming later: agent swarms and teams at scale, and microVM isolation. It is model-agnostic. If you choose a cloud model, your prompts go to that provider; your agents, files and secrets stay on the server. Add GPUs later and it becomes a TQ–192 base. Warranty: 3-year parts warranty, labour included, with carry-in service: bring or courier the server to us, and TensorQuay diagnoses, repairs and handles the manufacturer claims. Pickup and return is available as a paid option, or with the maintenance plan. Not covered: accidental damage, misuse, liquid damage or unauthorised modifications. Choose your quantity and pay by card or PayNow through Stripe checkout, with processing fees included. We then agree your team’s setup and the delivery date with you. What exactly do I receive? One configuration, supplied one way. An assembled workstation running Ubuntu, burn-in tested, with Qwen3.8-27B in NVFP4 served by vLLM and an AI agent configured for the workflows we agree with you. Plus documentation written around your setup and a guided handover. We do not sell the hardware without the AI stack installed, and there is no cheaper bare-metal option. The agent interface itself is still under evaluation, so we will tell you which one you are getting before you order. TQ–192 is available by enquiry only. Can I change the CPU, RAM or SSD? No. Each platform has a fixed specification, so we can qualify it once, document it once and support one repeatable setup rather than a different machine every time. Storage is 2 TB. You choose the platform, not the parts. If your work genuinely needs something outside the specification, tell us before ordering and we will say honestly whether we can support it. Have you fine-tuned the model weights? The current documented Qwen bundle uses serving and agent configuration tuning. It is not a claim that we have trained a new model checkpoint. Any customer-specific weight fine-tuning would need a separate agreed scope, dataset and evaluation. How should I read the performance numbers? They come from internal September 2026 tests on a rented RTX 5090 cloud system. The displayed throughput uses short scripted requests with up to 512 output tokens. Longer conversations, retrieval and tool-using agents change speed and memory use. The assembled workstations still need integration, thermal and extended reliability testing. Does everything work offline? The installed model can run inference on your workstation without a cloud API. A workflow that uses only local tools and data can work offline; internet requirements depend on the integrations you choose. Model downloads, software updates and optional tools such as web search, email or a cloud drive need a connection and may send data to an external service. We document the agreed data flow and network requirements at handover. The assistant on this website uses a cloud AI service and is separate from your workstation. How will my workstation be delivered? Scheduled delivery within Singapore. We arrange a delivery date with you after your order. Delivery only — we do not offer customer collection. Any delivery charge is confirmed as part of your order, not added afterwards. We also agree how to run your guided handover: on site with you, or remotely, whichever suits. Will my workstation show usage hours when it arrives? Yes, and that is deliberate. Your machine will show roughly 70–100 power-on hours and some disk writes when you receive it. It is not a used or refurbished system. Those hours are our work. Before a workstation leaves us we install and configure the operating system, drivers, the AI models and the agent, then run it under sustained load for around 72 hours — GPU stress, memory test, disk write and the inference stack at working concurrency. Most component failures happen in the first days of use, so we would rather find them here than at your desk. A machine showing zero hours has not been tested by anyone. This is the same reason a laptop bought with an operating system preinstalled already has hours on it: the software had to be installed and verified by someone. We plan to include the drive’s SMART report in your handover documents, showing exactly what was on the clock at dispatch and what the burn-in consisted of, alongside every component’s serial number and warranty expiry. You can verify it yourself at any time with sudo nvme smart-log /dev/nvme0n1 . What warranty and after-sales support do I get? Every TQ–32 comes with a 3-year parts and labour warranty on the workstation hardware . TensorQuay handles warranty claims: you contact us, we diagnose the fault and arrange the repair, so you deal with one company rather than several. Service is pickup and return in the first year, then carry-in service, not on-site repair. Repair time depends on parts availability, and a GPU replacement can take longer. The warranty does not cover accidental damage, misuse, liquid damage or unauthorised modifications. Ask us for the full warranty terms before ordering. We also plan to buy only stock carrying local Singapore warranty, never parallel imports, because a grey-market part has no local claim route and any promise made on it would be empty. Every deployment includes the agreed configuration and handover. Ongoing software support is the optional S$299/year maintenance plan; new AI workflows beyond the agreed scope are quoted separately. Do I have to pay for software maintenance? No. Software maintenance is optional at S$299 a year and you can start it at any time. Without it your machine keeps working exactly as delivered — nothing expires, nothing stops, and no feature is withheld. What it covers: model and runtime updates, driver and security patches, email support, and remote sessions that you start, which are time-limited and close automatically. There is no standing access to your machine and nothing runs in the background reporting to us. Response times and the number of remote sessions included are being confirmed. Agree the written scope with us before starting the maintenance plan. Are you an NVIDIA partner? We have contacted NVIDIA about partner onboarding. NVIDIA Partner Network membership and direct purchasing eligibility have not been established. NVIDIA product names describe the hardware we plan to evaluate. LET’S TALK LET’S BUILD WHAT’S NEXT. ## Your next chapter. Powered locally. AI Stations, deployment, training and support. Tell us what you have in mind. [Email our team](mailto:hello@tensorquay.com?subject=TensorQuay%20enquiry) Start in chat, or send your requirements to our team by email. [hello@tensorquay.com](mailto:hello@tensorquay.com)