Host TonicPremium Cloud Infrastructure

AI Inference

Run LLMs at scale

Dedicated NVIDIA clusters with sub-second API inference.

Ready in ~15 minutes Hourly or monthly billing Data stays on your server
GPU-powered LLM hosting
Inference
vLLM
248
tokens / sec
Llama 3.1 70B · batch 32
TTFT
92 ms
Context
128K
Included in every plan
Included
  • Llama · Mistral · Mixtral
  • vLLM · Ollama · TGI
  • OpenAI-compatible endpoint
  • LoRA & QLoRA fine-tuning
  • GPU autoscaling
  • Data stays on your server
Ready in ~15 min
No per-token fees
Open-source first

Llama · Mistral · Mixtral · Gemma · Qwen · DeepSeek · Phi-4.

vLLM runtime

Paged-attention batching + streaming for high throughput.

Fine-tuning

LoRA & QLoRA workflows with your own datasets.

Autoscale

Replicas that follow traffic — no cold-start tax.

Currency

Choose billing period

Neuron

Fine-tune & serve small open-source models (up to 14B).

৳22,900/mo

৳28,625/mo · save 20%

৳549,600 billed every 24 months

Choose plan
  • 1× RTX 4090 · 24 GB GDDR6X
  • 36-core Dual Xeon · 256 GB RAM
  • 240 GB SSD + 2 TB NVMe + 8 TB SATA
  • Llama 3 8B · Mistral 7B · Phi-4 14B
  • Ollama · vLLM · llama.cpp preinstalled
  • OpenAI-compatible /v1/chat/completions
  • 1 Gbps unmetered · 1 dedicated IPv4
MOST POPULAR

Synapse

Serve mid-size 27B–34B models with room to breathe.

৳40,900/mo

৳51,125/mo · save 20%

৳981,600 billed every 24 months

Choose plan
  • 1× RTX A6000 · 48 GB GDDR6
  • 36-core Dual Xeon · 256 GB RAM
  • 240 GB SSD + 2 TB NVMe + 8 TB SATA
  • Gemma 2 27B · DeepSeek-R1 14B · Mixtral 8×7B
  • vLLM auto-batching & streaming
  • LoRA / QLoRA fine-tuning workspace
  • Priority ML engineer support

Cortex

Production inference for 70B-class models on A100.

৳155,900/mo

৳194,875/mo · save 20%

৳3,741,600 billed every 24 months

Choose plan
  • 1× NVIDIA A100 · 80 GB HBM2e
  • 36-core Dual Xeon · 256 GB RAM
  • Multi-model serving · replica reuse
  • Llama 3.3 70B · Qwen 2.5 72B
  • Redis token cache · Prometheus metrics
  • Autoscaling replicas · JWT / OAuth2
  • 99.9% uptime SLA

Hypermind

Frontier H100 clusters with InfiniBand interconnect.

Custom

Tailored to your requirement — pay only for what you need.

  • H100 / H200 GPU clusters (1–8× per node)
  • InfiniBand interconnect · NVLink
  • Llama 3.1 405B · DeepSeek-V3 · custom weights
  • TensorRT-LLM · DeepSpeed · multi-GPU tensor parallel
  • Private endpoints · VPC peering · SOC 2
  • Dedicated MLOps team · 24/7 AI support
  • Enterprise SLA & compliance (HIPAA, GDPR)

GPU availability varies by region. 1 USD = 125 BDT · 1 EUR = 155 BDT. Hourly billing available on request.

Featured models

Top open-source LLMs, pre-installed.

Skip the download-and-configure ritual. Every server ships with your chosen model, runtime and API layer ready to serve traffic in about 15 minutes.

Llama 3.3 70B
by Meta

State-of-the-art 70B model — parity with Llama 3.1 405B for most tasks.

DeepSeek-R1 14B
by DeepSeek

First-gen reasoning model with performance comparable to OpenAI o1.

Gemma 2 27B
by Google

High-performance efficient model available in 2B, 9B and 27B variants.

Mixtral 8×7B
by Mistral AI

Sparse Mixture-of-Experts with strong reasoning at lower inference cost.

Phi-4 14B
by Microsoft

14B state-of-the-art open model tuned for reasoning and math.

Qwen 2.5 72B
by Alibaba

Long-context multilingual model with excellent code + tool use.

Right-sizing

Pick a GPU by model size.

A rough rule: VRAM ≥ model size × 1.2 for Ollama / TGI, or × 1.5 for vLLM. Multi-GPU nodes unlock tensor parallelism for the largest frontier models.

  • Ollama: ≥ Model × 1.2 VRAM
  • vLLM: ≥ Model × 1.5 VRAM · strong multi-GPU
  • TGI: ≥ Model × 1.2 · strong multi-GPU
  • DeepSpeed: ≥ Model × 1.1 · super-strong (NVLink)
GPU
Best for
Example models
RTX 4090 · 24 GB
≤14B params
Llama 3 8B · Mistral 7B · Phi-4
RTX A6000 · 48 GB
14B–34B
Gemma 2 27B · Mixtral 8×7B
A100 · 80 GB
70B (quantized) — 34B FP16
Llama 3.3 70B · Qwen 2.5 72B
H100 · 80 GB (cluster)
70B+ FP16 · 405B tensor-parallel
Llama 3.1 405B · DeepSeek-V3
LLM hosting architecture diagram

Architecture

Modular, production-ready inference stack.

Every deployment ships with the pieces you'd otherwise assemble by hand — load balancing, an OpenAI-compatible gateway, batched inference, Redis token cache and full observability.

  • Load balancer · API gateway
  • vLLM / TensorRT-LLM engine
  • Redis token cache
  • A100 / H100 GPU cluster
  • Model registry (S3 / NFS)
  • Prometheus + Grafana

Drop-in API

A drop-in replacement for the OpenAI API.

Point your existing SDK at your Host Tonic endpoint and keep your data, your weights, and your latency budget. No refactor required.

  • OpenAI /v1/chat/completions
  • Streaming & auto-batching
  • LoRA / QLoRA fine-tuning
  • Private model registry
  • Multi-model serving
  • Autoscaling replicas
  • Token & latency dashboards
  • API keys · OAuth2 · JWT
chat.py
from openai import OpenAI

client = OpenAI(
    base_url="https://llm.your-domain.com/v1",
    api_key="ht_live_••••••••",
)

resp = client.chat.completions.create(
    model="llama-3.3-70b",
    messages=[{"role": "user",
               "content": "Ship it."}],
    stream=True,
)

for chunk in resp:
    print(chunk.choices[0].delta.content or "", end="")

Why self-host

The case for a dedicated LLM server.

Data privacy & compliance

Full data residency, on-prem-style compliance (HIPAA, GDPR), controlled logs and audit trails.

Predictable latency

Dedicated GPUs, Redis token caching and load balancing — no noisy neighbours.

No per-token billing

Pay for the server; run as many tokens as it can produce. Zero surprise invoices.

Version stability

You control the model version. No silent upgrades, no behaviour drift.

Multi-GPU parallelism

Tensor + pipeline parallelism for 30B–70B+ models with NVLink / InfiniBand.

No vendor lock-in

Break free from API rate limits and cloud dependencies — your endpoint, forever.

Runtimes & tools

Every inference stack you'd reach for.

vLLM

Paged-attention batching · streaming · high throughput

Ollama

One-command model management · perfect for dev + prototypes

llama.cpp

GGUF quantization · CPU + GPU hybrid inference

TGI

Hugging Face Text-Generation-Inference · tensor parallel

TensorRT-LLM

NVIDIA-optimised kernels for H100 / A100

OpenWebUI

Beautiful chat UI on top of your private endpoint

Frequently asked questions

How fast can I deploy an LLM server?+

Around 15 minutes. Our auto-provisioner installs the GPU driver stack, your chosen runtime (vLLM / Ollama / TGI) and pre-pulls the model weights so you can hit /v1/chat/completions on first boot.

Which GPUs are available?+

RTX 4090 (24 GB), RTX A6000 (48 GB), A100 (80 GB HBM2e) and H100 (80 GB) — single or multi-GPU nodes with NVLink / InfiniBand for the larger clusters.

Is the endpoint really OpenAI-compatible?+

Yes. /v1/chat/completions, /v1/completions and /v1/embeddings match the OpenAI schema, including streaming and function calling. Point any OpenAI SDK at your endpoint.

Do I pay per token?+

No. You rent the server; run as many tokens as the hardware can produce. Billing is either monthly or hourly — pick whichever fits your workload.

Can I fine-tune models on my own data?+

Yes — Neuron and above ship with a LoRA / QLoRA workspace. Bring a JSONL dataset, kick off a run, and load the adapter into vLLM without redeploying.

Where does my data live?+

On the server you rent. Weights, prompts and completions never leave your instance. Storage is encrypted at rest and the network is protected end-to-end.