deployment● live
model acme/companion-v4
capacity dedicated
region us-central
tier priority
Dedicated deployments
Reserved capacity and your own model ID, with no queueing behind other customers.
Dedicated inference built for people
Dedicated capacity for high-volume realtime workloads, with no queueing behind other customers.
Tens of trillions of tokens a month
What a dedicated deployment includes
We host your weights, fine-tune on your requests and optimize decoding and kernels for them.
"model": "you/companion-v4",
"service_tier": "priority"
"service_tier": "priority"
✓ weights loaded · fp8 · 31B
Bring your own model
Upload weights or point us at a repo, and your model runs on the same engine and API as ours.
fine-tuning run
Fine-tuning on your traffic
Fine-tune open models on your production conversations.
ahead the knees need a warm…
check ✓ ✓ ✓ ✓ ✕
accept 4 tokens in 1 pass
check ✓ ✓ ✓ ✓ ✕
accept 4 tokens in 1 pass
Drafts tuned on your traffic
Draft models trained on your requests.
realtime decoding · verify attention
up to 7.5×
faster at checking drafted tokens, per attention layer
Gemma 4 26B · microbenchmark
Kernels and engines
When a stock engine is the limit, we optimize the attention step, the KV-cache router or the serving backend for your model.
regions
● us-central
● eu-west
● asia-south
● asia-east
● us-central
● eu-west
● asia-south
● asia-east
zero data retention
on
on
Regions and data
US, Europe, India or Asia Pacific, with zero data retention on the models we host.
The fleet your deployment runs on
Dedicated deployments run on the fleet we operate for every Inworld model, so a fix we make reaches your traffic in days.
Tens of trillions of tokens a month
Billionsof requests a month
Peak above 1,100 a second
Millionsof tokens a second
Around the clock
Global infrastructure
Four live regions
- TTS
- LLM
- STT
- Speech to speech
Voice across four regions. LLM inference in the US.
Illustrative activityUp to 2.8×faster first token
Gemma 4 31B, median, 943 → 336 ms
Up to 10×faster first token for the slowest 1%
DeepSeek V4 Flash, cache-aware routing, 9.4 → 0.87 s
Up to 90%cache hits on repeated prompts
Gemma 4 26B
99.999%of requests succeeded
More than half a billion Gemma 4 31B requests in 30 days, about 1 in 160,000 failed
Each figure is the best measured case on the model it names.
Inference built for people runs on Inworld. Who are you building for?
Questions
What is a dedicated LLM deployment?
A dedicated deployment on Inworld AI is inference capacity reserved for your traffic on the fleet we operate, with your own model ID. Your requests never queue behind other customers, and your model stays loaded between requests. The same fleet serves tens of trillions of tokens a month, and the team that runs it tunes your deployment.
Can I bring my own model?
Yes, you can bring your own open or fine-tuned model to Inworld AI for dedicated inference, by uploading weights or pointing us at a repo. We select the inference stack and tune quantization, caching, batching and parallelism for your workload, then deploy it. If you want to run your own serving stack, we also offer Kubernetes access to dedicated nodes.
How do I call a dedicated model?
You call a model on an Inworld AI dedicated deployment the same way you call every Inworld model, with Chat Completions requests to https://api.inworld.ai/v1/chat/completions, your Inworld API key and your model ID. The OpenAI SDK works with its base URL set to https://api.inworld.ai/v1, so moving traffic from a shared model to your deployment means changing one model ID.
Can I run Realtime TTS-2 on dedicated capacity?
Yes, Inworld AI's voice models, including Realtime TTS-2 text to speech, run on dedicated capacity too, with guaranteed capacity and no shared tenancy. A dedicated voice deployment suits apps with steady high volume that want capacity reserved for their own traffic. Realtime TTS-2 Flash returns first audio in as little as 50 ms.
Can you fine-tune a model on my traffic?
Yes, fine-tuning on your production conversations is available on an Inworld AI dedicated deployment, with your approval. We fine-tune open models on your traffic, and we can train draft models on your requests for realtime decoding, which drafts several tokens ahead and has the full model verify them in one pass.
Where can a dedicated deployment run, and is zero data retention available?
An Inworld AI dedicated deployment can run in the US, Europe, India or Asia Pacific, the four regions of the fleet we operate. Zero data retention is available on Enterprise plans and as an add-on on the Growth plan for the models we host, so prompts and completions are processed in memory and redacted from our logs as soon as processing completes. Usage metadata is kept for billing.
How is dedicated capacity priced?
Dedicated inference capacity on Inworld AI is priced on demand or provisioned, with volume pricing and no egress fees. Each quote is set for your workload and traffic, and capacity is reserved for you, so your requests never queue behind other customers. For comparison, shared serving of open models starts from $0.07 per million input tokens, for Gemma 4 26B.
Is dedicated inference ready for production?
Yes, Inworld AI dedicated deployments run production LLM traffic with SOC 2 Type II controls, isolated tenancy and monitoring around the clock with automatic recovery. On the same fleet, 99.999% of requests succeeded on Gemma 4 31B over 30 days, more than half a billion requests with about 1 in 160,000 failing, counting server errors, timeouts and capacity rejections.