One endpoint.GPUs that scale to zero.
Call open-weight chat and embedding models through an OpenAI-compatible API. Tessellate starts GPUs when requests arrive and stops them when they don't.
Features
Run models without running GPUs
Autoscale to zero
GPUs start when a request arrives and shut down when traffic stops, so idle models cost you nothing.
Pay per token
You are billed for the tokens you send and receive, not for reserved GPUs or hours of uptime.
OpenAI-compatible API
Point your existing OpenAI client at Tessellate by changing the base URL and key. No new SDK to learn.
Wire up an endpoint in a minute
Connect a model, hardware, regions and a prompt. Kestrel turns the graph into a live API.
- Model weights
- Compute
- Routing
- System prompt
Support widget
Waiting
Slack bot
Waiting
Eval suite
Waiting
Usage alerts
Waiting
Model
llama-3.1-70b-instruct
fp8 · 128k context
Hardware
8 × H100 80GB
Autoscale 2–64 replicas
Regions
System prompt
Answer from the help center. Link the page you used.
Endpoint
support-agent- Model weights
- Compute
- Routing
- System prompt
Awaiting inputs · 1/4
Support widget
help.northwind.dev
Chat on every page
Slack bot
#support-escalations
Replies in threads
Eval suite
1,200 golden prompts
Runs on every deploy
Usage alerts
$2,000 monthly budget
Email and PagerDuty
Drag from a dot to the matching input. 1 of 4 connected.
Requests / sec
48,210
Regions live
7
Virginia, us-east, 38 milliseconds median latency.
One endpoint, seven regions
Deploy once. Kestrel serves your model from every region on the map.
Nearest-region routing
Each request goes to the closest healthy region. You change no DNS.
Automatic failover
If a region slows down, traffic moves to the next nearest one in seconds.
Health checks every 10 sec
We test every GPU node and drain bad hardware before it serves a token.
Developers
Change the base URL. Keep your client.
Chat completions and embeddings use the same request shapes as the OpenAI API. Set your key in TESSELLATE_API_KEY and send the request.
curl https://api.tessellate.example/v1/chat/completions \
-H "Authorization: Bearer $TESSELLATE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "llama-3.3-70b",
"messages": [{"role": "user", "content": "Explain autoscaling to zero."}]
}'The host in these samples is illustrative. Your key page lists the endpoint for your account.
Models
Four open-weight models, one endpoint
Three chat models and one embedding model, all served through the same OpenAI-compatible API.
- Chat
Llama 3.3 70B
Meta
General-purpose chat and instruction following for assistants and agents.
- Chat
Qwen 2.5 32B
Alibaba Cloud
Multilingual chat with strong code and structured-output handling.
- Chat
Mistral Small 3
Mistral AI
A compact chat model for high-volume, everyday tasks.
- BGEEmbedding
BGE-M3
BAAI
Multilingual embeddings for search, retrieval and clustering.
Pricing
Start free, pay for the tokens you serve, and reserve GPUs when your traffic is steady.
Hobby
For side projects and trying out models.
Free
No credit card needed
Includes:
- $5 in free credits each month
- Serverless endpoints
- 60 requests per minute
- Community support
Pro
Save 20%For teams that ship their first AI features.
$39
per workspace per month, billed yearly
Everything in Hobby, plus:
- $25 in monthly credits
- Every model in the catalog
- 600 requests per minute
- LoRA fine-tuning
- All 7 regions
Dedicated
Save 20%For steady traffic that needs fixed capacity.
$1.99
per H100 hour, 12-month term
Everything in Pro, plus:
- Reserved H100 and H200 GPUs
- No rate limits
- Autoscaling floors and caps
- Private model weights
- Priority support
$500 in creditsIncluded
Reserve GPUsEnterprise
For large teams with security and compliance needs.
Custom
Volume pricing
Everything in Dedicated, plus:
- Private regions and VPC peering
- SSO and audit logs
- 99.99% uptime SLA
- A named support engineer
- Invoicing and custom terms
| Features | HobbyFree | Pro$39 per workspace per month | Dedicated$1.99 per H100 hour | EnterpriseCustom |
|---|---|---|---|---|
| Usage | ||||
| Monthly credits | $5 | $25 | $500 | Custom |
| Requests per minute | 60 | 600 | No limit | No limit |
| Concurrent requests | 5 | 50 | 500 | Custom |
| Batch API | Not included | Included | Included | Included |
| Models | ||||
| Catalog models | Top 10 | All | All | All |
| LoRA fine-tuning | Not included | Included | Included | Included |
| Upload your own weights | Not included | Not included | Included | Included |
| Pinned model versions | Not included | Included | Included | Included |
| Infrastructure | ||||
| Regions | 3 | 7 | 7 | 7 + private |
| Reserved GPUs | Not included | Not included | Included | Included |
| Cold start | < 10 s | < 2 s | None | None |
| VPC peering | Not included | Not included | Not included | Included |
| Security and support | ||||
| SSO and audit logs | Not included | Not included | Not included | Included |
| SOC 2 Type II report | Not included | Included | Included | Included |
| Uptime SLA | Not included | 99.9% | 99.95% | 99.99% |
| Support | Community | Priority | Named engineer | |
Send your first request today
Create a key, point your OpenAI client at Tessellate, and call a model.
Get an API key