LLM_Deploy
Talk to an engineerMenu

You host it.You own it.

$4,900 · 72 hours · fixed-price pilot

Open-weight LLMs deployed inside your network, on GPUs you control, behind an OpenAI-compatible endpoint. Run by the team that has kept enterprise systems online since 2000. Prove it on your data in a fixed-price 72-hour pilot.

Start the $4,900 pilotTalk to an engineer

An UpSystems practice — enterprise infrastructure since 2000.

Your network, your endpoint.

Your applications keep the OpenAI SDK they already use. Only the base URL changes, and it points at a host inside your network.

Python · drop-inOpenAI-compatible
client = OpenAI(
    base_url="https://llm.yourcompany.internal/v1",   # inside your network
    api_key="your-key",
)

Same SDK. Change the base URL. Tested-integration matrix on request.

Your models, sized to your GPUs.

We choose the model and the GPUs together, then tune quantization, batching and caching until it fits the hardware you have.

Llama 4 Scout
1× H100 80GB at 4-bit, 1× H200 141GB at FP8
Qwen3
CPU only at 4-bit, 1× T4 16GB at FP8 (8B)
gpt-oss-20b
1× T4 16GB at 4-bit, 1× RTX 5090 32GB at FP8
DeepSeek V3 / R1
8× H100 (640GB) at 4-bit, 8× H200 (1,128GB) at FP8
All 60 models and their footprints

Operated 24/7 from Limassol and Auckland.

Model and driver updates, monitoring and on-call stay with UpSystems, the team that has kept enterprise systems online since 2000.

99.95%
uptime SLA
5 min
incident response
22
engineers
100+
clients

Three ways in.

External APIExternal APIExternal APIPrompts leaveYour networkprompts + data
Cloud APIDEMO TELEMETRY
Metered spend~$10,000per month, as reported
100%
Egress> 0prompts leave your network
Last response429RATE LIMITED
External endpoints1+reached per prompt

Cloud APIor your own GPUs.

Your networkClientEndpointGPU
Your own GPUsDEMO TELEMETRY
Run cost~$2,000per month, as reported
20%
Egress0 Bnothing leaves
Last response200OK
External endpoints0reached

A hosted API meters every prompt and sends it to the provider's servers. Open-weight models on your own GPUs answer inside your network. Drag the divider to see both. Cost figures are as reported by customers; the telemetry is a demo.

Same prompts. A different bill, a different perimeter.

Rows marked reported are customer figures from installations we run. They are not a benchmark, and your workload will differ.

Cloud API compared with your own GPUs, row by row
MeasureCloud APIYour own GPUs with LLM Deploy
Data flowPrompts and outputs travel to the provider’s servers.Prompts and outputs stay inside your network.
Monthly billNocodo LTD, reported~$10,000
~$2,000
240 hackathon submissions scoredInfiano, reported≈$6,750
≈$3,000
LimitsRate limits and quotas are set by the provider.Throughput is what the GPUs you size can serve.
LatencyNocodo LTD, reported1×
1×
3× faster
3×
1.1–1.4 s to first token on the 40-GPU Infiano cluster, 0 restarts.
Air-gapNeeds a route out to the provider’s endpoint.The endpoint sits inside your network, so inference needs no route out.

A hosted API is the better call when volume is low or spiky, you have no GPUs to run, and nothing in the prompts is sensitive. Owning the stack pays off as volume and data sensitivity grow.

Numbers from installations we run.

Two deployments and the team behind them, itemised.

Nocodo LTD

Monthly model billBefore / after

OpenAI API~$10,000
Own infrastructure~$2,000
Response speed3× faster
Compliance overheaddown ~60%
Difference, per month~$8,000

“LLMDeploy team has unique expertise in understanding our situation and adapting to our needs, while delivering the fastest, reasonable and very affordable winning solutions.”

Almas Abulkhairov, CEO, Nocodo LTD

Infiano hackathon

Event inference billCloud vs on-prem

Cloud API≈$6,750
40-GPU cluster≈$3,000
Cluster stood up in48 hours
Submissions scored240
Time to first token1.1–1.4 s
Infrastructure-caused restarts0
Difference, event≈$3,750

UpSystems

The team behind itSince 2000

Engineers22
Clients100+
OfficesLimassol · Auckland
Coverage24/7
Uptime SLA99.95%
Incident response5 min

From a 72-hour pilot to 24/7 operations.

Each step is priced up front. The pilot readout scopes production, and production hands over to managed operations.

  1. Pilot

    $4,900fixed

    • One open-weight model live on your data in 72 hours. The clock starts after access and prerequisites.
    • OpenAI-compatible endpoint.
    • Reference evaluation on your data.
    • Readout and production scoping.
    1. 00 hours: Access
    2. 08 hours: Model pulled
    3. 20 hours: Quantized
    4. 36 hours: Endpoint up
    5. 54 hours: Evaluation
    6. 72 hours: Readout
    Sample schedule access, model pulled, quantized, endpoint up, evaluation, readout. At 72 you keep: measured result on your data, scoped plan for production.
    Start the $4,900 pilotPrerequisites and pilot terms

  2. Production deployment

    from $25,000one-time, hardware separate

    • Model selection and GPU sizing.
    • Quantization, batching, caching.
    • Monitoring, RBAC, audit logging.
    • OpenAI-compatible API inside your network.
    • Handover and documentation.

    We size the hardware and source it at cost, with no markup. One H100 80GB ≈ $25–30k, a 4×H100 node ≈ $180–240k; smaller models run on far less.

  3. Managed ops

    from $1,500per GPU node / month

    • Model and driver updates, monitoring, on-call, SLA tiers.
    • 99.95% uptime SLA and 5-minute incident response.
    • 24/7 across Limassol and Auckland.

How it is set up, and who runs it.

Your GPUs, the model and the endpoint sit inside your network. UpSystems can run it for you 24/7, and confidential compute covers GPUs you do not own.

01 / 03 · Your installation

Inside your network.

Open-weight LLMs deployed inside your network, on GPUs you control, behind an OpenAI-compatible endpoint.

  • Your GPUs, sized to the model and sourced at cost.
  • The model, with quantization, batching and caching tuned to fit them.
  • OpenAI-compatible API inside your network.
  • RBAC and an audit log on the endpoint.

from $25,000one-time, hardware separate

02 / 03 · Managed ops, 24/7

We run it for you.

Run by the team that has kept enterprise systems online since 2000, from Limassol and Auckland. Health metrics reach the on-call team. Prompts, completions and logs stay on your GPUs and in your storage.

  • Model and driver updates, monitoring, on-call, SLA tiers.
  • 99.95% uptime SLA and 5-minute incident response.
  • 24/7 across Limassol and Auckland.

from $1,500per GPU node / month

03 / 03 · Confidential compute

On GPUs you do not own.

Confidential computing changes whom you have to trust. You trust the CPU and GPU hardware, and the attestation that proves it was switched on.

  • Your prompts and the answers. They are decrypted only inside the enclave and at your endpoint.
  • Your model weights and fine-tunes. They load into the GPU’s protected memory region, which the host cannot read.
  • The host can still see that a workload is running, the size and timing of the traffic, and the attestation record.
  • NVIDIA H100, H200 and Blackwell, running in confidential-computing mode.

By quote

Type a prompt. See who can read it.

The same request, seen from both ends. You get a plain answer. The company that owns the GPU sees encrypted memory traffic and an attestation record, and nothing it can read.

Your view

Sample

> Summarise the open claims for policy 4471-B

Sample response. This page runs no model, so this is a canned reply. In production your model answers inside the sealed enclave, and the reply is readable only at your end of the encrypted session.

Plain text, at your endpoint only.

Attestation

Demo
  1. CPU TEE report verifiedPass
  2. GPU in confidential-computing modePass
  3. GPU attestation report signed by NVIDIAPass
  4. Driver and firmware measurements matchPass
  5. CPU to GPU transfers encryptedPass
  6. Keys released to the enclave onlyPass

6 of 6 checks passed

Hardware owner's view

Sample

Unreadable encrypted bytes. The hardware owner cannot see your prompt or the answer.

Readable by the host: nothing above.

Illustration only. No model runs on this page and nothing leaves your browser. The bytes on the right are a SHA-256 stream of your prompt, so they change with every keystroke, the way real ciphertext would.

Open-weight models we deploy.

The 60 most-downloaded open-weight models for chat, code, vision, OCR, search and speech. Pick your GPU and the ones that fit move to the top.

Category
Your GPU

8 of 60 models · by popularity

Open-weight models we deploy: model and vendor, category, total and active parameters, context, estimated VRAM at FP8 and 4-bit, licence, and the smallest GPU configuration it fits on. Ranked by Hugging Face downloads.
ModelCategorySize / activeContextest. VRAM FP8LicenceFits on
Gemma 4 31BGoogleChat + vision31B dense256K37.5 GB4-bit 18.8Apache 2.01× RTX 4090 24GB4-bitFP8 · 1× A100 40GB
Qwen3-Coder-30B-A3BAlibabaCoding30.5B MoE3.3B active256K36.6 GB4-bit 18.3Apache 2.01× RTX 4090 24GB4-bitFP8 · 1× A100 40GB
Unlimited-OCRBaiduOCR3.3B MoE~0.57B active32K4 GB4-bit 2MITCPU only4-bitor any GPU
all-MiniLM-L6-v2sentence-transformersEmbedding22.7M512 tokens<0.1 GB4-bit <0.1Apache 2.0CPU only4-bitor any GPU
Whisper large-v3-turboOpenAISpeech-to-text809M30 s windows1 GB4-bit 0.5MITCPU only4-bitor any GPU
Qwen3AlibabaChat8B / 32B / 235B-A22B22B (235B) active32K (128K with YaRN)9.8–282 GBacross sizes · 4-bit from 4.9Apache 2.0CPU only4-bitFP8 · 1× T4 16GB
Qwen3-Coder-NextAlibabaCoding80B MoE3B active256K95.6 GB4-bit 47.8Apache 2.01× L40S 48GB4-bitFP8 · 1× RTX PRO 6000 96GB
Gemma 4 26B A4BGoogleChat + vision26B MoE~4B active256K31 GB4-bit 15.5Apache 2.01× T4 16GB4-bitFP8 · 1× RTX 5090 32GB
DeepSeek-OCRDeepSeekOCR3.3B MoE~0.57B active8K4 GB4-bit 2MITCPU only4-bitor any GPU
BGE-M3BAAIEmbedding568M8K tokens0.7 GB4-bit 0.3MITCPU only4-bitor any GPU
Whisper large-v3OpenAISpeech-to-text1.55B30 s windows1.9 GB4-bit 0.9Apache 2.0CPU only4-bitor any GPU
Devstral Small 2Mistral AICoding24B dense256K28.8 GB4-bit 14.4Apache 2.01× T4 16GB4-bitFP8 · 1× RTX 5090 32GB
Qwen3.8-27BAlibabaChat + vision27B dense256K (1M ext.)33.3 GB4-bit 16.7Apache 2.01× RTX 4090 24GB4-bitFP8 · 1× A100 40GB
GLM-OCRZ.aiOCR0.9B128K1.6 GB4-bit 0.8MITCPU only4-bitor any GPU
bge-reranker-v2-m3BAAIReranker568M8K tokens0.7 GB4-bit 0.3Apache 2.0CPU only4-bitor any GPU
Voxtral Mini 4B RealtimeMistral AISpeech-to-text4.4Bstreaming5.3 GB4-bit 2.7Apache 2.0CPU only4-bitor any GPU
Kimi K2.7 Code on Hugging Face (opens in a new tab)Moonshot AICoding1T MoE32B active256K1,232 GB4-bit 616Modified MITattribution8× H1004-bitFP8 · multi-node
GLM-5.3-Flash on Hugging Face (opens in a new tab)Z.aiChat + vision320B MoE18B active1M386 GB4-bit 193MIT4× H1004-bitFP8 · 8× H100
LightOnOCR-2-1B on Hugging Face (opens in a new tab)LightOnOCR1B16K1.2 GB4-bit 0.6Apache 2.0CPU only4-bitor any GPU
Qwen3-Embedding-0.6B on Hugging Face (opens in a new tab)AlibabaEmbedding0.6B32K0.7 GB4-bit 0.4Apache 2.0CPU only4-bitor any GPU
Qwen3-ASR-1.7B on Hugging Face (opens in a new tab)AlibabaSpeech-to-text2.3Baudio, 64K2.8 GB4-bit 1.4Apache 2.0CPU only4-bitor any GPU
Qwen3-Coder-480B-A35BAlibabaCoding480B MoE35B active256K576 GB4-bit 288Apache 2.04× H1004-bitFP8 · 8× H100
Qwen3.5-9B on Hugging Face (opens in a new tab)AlibabaVision9B dense256K (1M ext.)11.6 GB4-bit 5.8Apache 2.0CPU only4-bitFP8 · 1× T4 16GB
DeepSeek-OCR 2DeepSeekOCR3.4B MoE~0.57B active8K4.1 GB4-bit 2Apache 2.0CPU only4-bitor any GPU
Qwen3-Reranker-4B on Hugging Face (opens in a new tab)AlibabaReranker4B32K4.8 GB4-bit 2.4Apache 2.0CPU only4-bitor any GPU
Nemotron 3.5 ASR Streaming 0.6B on Hugging Face (opens in a new tab)NVIDIASpeech-to-text638Mstreaming0.8 GB4-bit 0.4OpenMDW-1.1CPU only4-bitor any GPU
gpt-oss-20bOpenAIChat~21B MoE3.6B active128K25.1 GB4-bit 12.5Apache 2.01× T4 16GB4-bitFP8 · 1× RTX 5090 32GB
North Mini Code 1.0 on Hugging Face (opens in a new tab)CohereCoding30B MoE3B active256K36.6 GB4-bit 18.3Apache 2.01× RTX 4090 24GB4-bitFP8 · 1× A100 40GB
Qwen3-VL-8B-Instruct on Hugging Face (opens in a new tab)AlibabaVision8B dense256K (1M ext.)10.5 GB4-bit 5.3Apache 2.0CPU only4-bitFP8 · 1× T4 16GB
dots.ocr on Hugging Face (opens in a new tab)rednoteOCR3B128K3.6 GB4-bit 1.8MITCPU only4-bitor any GPU
Qwen3-Embedding-8B on Hugging Face (opens in a new tab)AlibabaEmbedding8B32K9.1 GB4-bit 4.5Apache 2.0CPU only4-bitFP8 · 1× T4 16GB
Parakeet TDT 0.6B v3 on Hugging Face (opens in a new tab)NVIDIASpeech-to-text627Mlong-form0.8 GB4-bit 0.4CC-BY-4.0attributionCPU only4-bitor any GPU
Qwen3.6-35B-A3B on Hugging Face (opens in a new tab)AlibabaChat35B MoE3B active256K (1M ext.)43.1 GB4-bit 21.6Apache 2.01× RTX 4090 24GB4-bitFP8 · 1× L40S 48GB
Qwen3-VL-235B-A22B-Instruct on Hugging Face (opens in a new tab)AlibabaVision235B MoE22B active256K (1M ext.)283 GB4-bit 141Apache 2.02× H1004-bitFP8 · 4× H100
olmOCR-2-7B-1025 on Hugging Face (opens in a new tab)Allen AIOCR8.3B128K9.9 GB4-bit 5Apache 2.0CPU only4-bitFP8 · 1× T4 16GB
EmbeddingGemma 2 on Hugging Face (opens in a new tab)GoogleEmbedding740M8K0.9 GB4-bit 0.4Apache 2.0CPU only4-bitor any GPU
DeepSeek-V4-Flash on Hugging Face (opens in a new tab)DeepSeekChat284B MoE13B active1M365 GB4-bit 183MIT4× H1004-bitFP8 · 8× H100
MiniCPM-V 4.7 35B-A3B on Hugging Face (opens in a new tab)OpenBMBVision35B MoE3B active256K42.3 GB4-bit 21.1Apache 2.01× RTX 4090 24GB4-bitFP8 · 1× L40S 48GB
PaddleOCR-VL-1.6 on Hugging Face (opens in a new tab)BaiduOCR0.9B128K1.2 GB4-bit 0.6Apache 2.0CPU only4-bitor any GPU
Nemotron 3.5 Rerank 8B on Hugging Face (opens in a new tab)NVIDIAReranker8.4B256K10.1 GB4-bit 5OpenMDW-1.1CPU only4-bitFP8 · 1× T4 16GB
Qwen3.8-Flash-NextAlibabaChat180B MoE6B active256K (1M ext.)216 GB4-bit 108Qwen Community License 1.0revenue cap1× H200 141GB4-bitFP8 · 4× H100
gpt-oss-120bOpenAIChat~117B MoE5.1B active128K140 GB4-bit 70.1Apache 2.01× H100 80GB4-bitFP8 · 1× H200 141GB
GLM-5.3 on Hugging Face (opens in a new tab)Z.aiChat753B MoE40B active1M904 GB4-bit 452GLM-5.3 License (MIT-style)conditions8× H1004-bitFP8 · 8× H200
DeepSeek-V4.1-Flash on Hugging Face (opens in a new tab)DeepSeekChat763B MoE8B / 16B active1M916 GB4-bit 458MIT8× H1004-bitFP8 · 8× H200
DeepSeek V3 / R1DeepSeekChat671B MoE37B active128K821 GB4-bit 411MIT (R1); DeepSeek Model Agreement (V3)use policy (V3)8× H1004-bitFP8 · 8× H200
Nemotron 3 Super 120B-A12B on Hugging Face (opens in a new tab)NVIDIAChat120B hybrid MoE12B active256K (up to 1M)148 GB4-bit 74.2NVIDIA Nemotron Open Model License1× H100 80GB4-bitFP8 · 2× H100
Hy3 on Hugging Face (opens in a new tab)TencentChat295B MoE21B active256K359 GB4-bit 179Apache 2.01× B200 180GB4-bitFP8 · 8× H100
Kimi K3 on Hugging Face (opens in a new tab)Moonshot AIChat2.8T MoE104B active1M3,336 GB4-bit 1,668Kimi K3 Licenserevenue capMulti-nodebeyond 8× H200
DeepSeek-V4-Pro on Hugging Face (opens in a new tab)DeepSeekChat1.6T MoE49B active1M1,919 GB4-bit 959MIT8× H2004-bitFP8 · multi-node
Gemma 3GoogleChat270M–27B dense128K32.9 GB4-bit 16.5Gemma Terms of Useuse policy1× RTX 4090 24GB4-bitFP8 · 1× A100 40GB
Mistral Small 3.1Mistral AIChat24B dense128K28.8 GB4-bit 14.4Apache 2.01× T4 16GB4-bitFP8 · 1× RTX 5090 32GB
MiniMax-M3 on Hugging Face (opens in a new tab)MiniMaxChat~428B MoE~23B active1M512 GB4-bit 256MiniMax Community Licenserevenue cap4× H1004-bitFP8 · 8× H100
Llama 4 ScoutMetaChat109B MoE17B active10M (claimed)130 GB4-bit 65.2Llama 4 Community Licenseconditions1× H100 80GB4-bitFP8 · 1× H200 141GB
Mistral Small 4Mistral AIChat119B MoE6.5B active256K143 GB4-bit 71.6Apache 2.01× H100 80GB4-bitFP8 · 2× H100
Command A+CohereChat218B MoE25B active128K263 GB4-bit 131Apache 2.01× H200 141GB4-bitFP8 · 4× H100
Llama 4 MaverickMetaChat400B MoE17B active1M482 GB4-bit 241Llama 4 Community Licenseconditions4× H1004-bitFP8 · 8× H100
Mistral Large 2Mistral AIChat123B dense128K147 GB4-bit 73.6Mistral Research License (MRL)research onlycommercial option: Mistral Large 31× H100 80GB4-bitFP8 · 2× H100
Muse Spark Open on Hugging Face (opens in a new tab)MetaChat816B MoE42B active1M986 GB4-bit 493Muse Model Licenselimited commercial8× H1004-bitFP8 · 8× H200
Mistral Large 3Mistral AIChat675B MoE41B active256K810 GB4-bit 405Apache 2.08× H1004-bitFP8 · 8× H200
Command ACohereChat111B dense256K133 GB4-bit 66.6CC-BY-NC-4.0non-commercialcommercial option: Command A+1× H100 80GB4-bitFP8 · 1× H200 141GB
> show all 60 models> show the top 8

est. VRAM = parameters × bytes per weight (BF16 2, FP8 1, 4-bit 0.5) × 1.2 for runtime overhead. Long-context KV cache comes on top. A configuration fits when the estimate is at or below its GPU memory. Estimates, not measurements. Rank within a category by trailing 30-day Hugging Face downloads of the primary repo, read 2026-10-27; embeddings and rerankers are ranked together. We size the real configuration in the pilot.

Operated from two offices, around the clock.

in enterprise IT
Since 2000
engineers
22
Cyprus and New Zealand
Limassol and Auckland
cover, 5 min incident response
24/7

Limassol

08:00 local

Auckland

19:00 local
Demo telemetry
06:00UTC

Chapters.

  1. /how-it-worksOne installation, five parts, inside your network
  2. /models60 open-weight models, sized to your GPUs60 models
  3. /industriesSame installation, your sector’s workflows4 sectors
  4. /securityNothing leaves; everything is logged
  5. /confidential-computeAttested confidential GPUs; what the host can and cannot see
  6. /compareSame endpoint, your GPUs, flat cost~$10,000 → ~$2,000/mo
  7. /pricingThree prices, published$4,900 · $25,000 · $1,500
  8. /pricing/calculatorYour numbers, not ours
  9. /pilotOne model live on your data in 72 hours, fixed price$4,900 · 72 h
  10. /production-deploymentYour installation, production-gradefrom $25,000
  11. /managed-opsTwo offices, one on-callfrom $1,500 per node / mo
  12. /case-studiesTwo deployments, itemised40 GPUs in 48 h
  13. /field-notesNotes from the operators7 articles
  14. /aboutEnterprise infrastructure since 2000, two offices22 engineers
  15. /contactAn engineer from Limassol or Auckland replies

An engineer from Limassol or Auckland replies.

Talk to an engineer.

Tell us the model, the data and the GPUs you have. An engineer from Limassol or Auckland scopes the 72-hour pilot with you.

Start the $4,900 pilotTalk to an engineer

Limassol HQ

Cyprus, 24/7

+357 25 123877

Auckland

New Zealand, 24/7

+64 9 887 9088